Skip AppImage and RPM builds to avoid slow 360MB+ compression

AppImage bundler compresses the entire sidecar.zip into squashfs,
causing builds to hang/timeout. Limit targets to deb (Linux),
nsis+msi (Windows), and dmg (macOS).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

.claude/settings.local.json

@@ -0,0 +1,21 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(git init:*)",
+      "Bash(git:*)",
+      "WebSearch",
+      "Bash(npm create:*)",
+      "Bash(cp:*)",
+      "Bash(npm install:*)",
+      "Bash(/home/jknapp/.cargo/bin/cargo test:*)",
+      "Bash(ruff:*)",
+      "Bash(npm run:*)",
+      "Bash(npx svelte-check:*)",
+      "Bash(pip install:*)",
+      "Bash(python3:*)",
+      "Bash(/home/jknapp/.cargo/bin/cargo check:*)",
+      "Bash(cargo check:*)",
+      "Bash(npm ls:*)"
+    ]
+  }
+}

.gitea/workflows/build-linux.yml

@@ -0,0 +1,124 @@
+name: Build Linux
+
+on:
+  push:
+    branches: [main]
+    tags: ["v*"]
+  pull_request:
+    branches: [main]
+
+env:
+  PYTHON_VERSION: "3.11"
+  NODE_VERSION: "20"
+  TARGET: x86_64-unknown-linux-gnu
+
+jobs:
+  build:
+    name: Build (Linux)
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      # ── Python sidecar ──
+      - name: Install uv
+        run: |
+          if command -v uv &> /dev/null; then
+            echo "uv already installed: $(uv --version)"
+          else
+            curl -LsSf https://astral.sh/uv/install.sh | sh
+            echo "$HOME/.local/bin" >> $GITHUB_PATH
+          fi
+
+      - name: Install ffmpeg
+        run: sudo apt-get update && sudo apt-get install -y ffmpeg
+
+      - name: Set up Python
+        run: uv python install ${{ env.PYTHON_VERSION }}
+
+      - name: Build sidecar
+        working-directory: python
+        run: uv run --python ${{ env.PYTHON_VERSION }} python build_sidecar.py --cpu-only
+
+      - name: Package sidecar for Tauri
+        run: |
+          cd python/dist/voice-to-notes-sidecar && zip -r ../../../src-tauri/sidecar.zip .
+
+      # ── Tauri app ──
+      - name: Set up Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: ${{ env.NODE_VERSION }}
+
+      - name: Install Rust stable
+        run: |
+          curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --default-toolchain stable
+          echo "$HOME/.cargo/bin" >> $GITHUB_PATH
+
+      - name: Install system dependencies
+        run: |
+          sudo apt-get install -y libgtk-3-dev libwebkit2gtk-4.1-dev libappindicator3-dev librsvg2-dev patchelf xdg-utils
+
+      - name: Install npm dependencies
+        run: npm ci
+
+      - name: Build Tauri app
+        run: npm run tauri build
+
+      # ── Release ──
+      - name: Upload to release
+        if: github.ref == 'refs/heads/main' || startsWith(github.ref, 'refs/tags/v')
+        env:
+          BUILD_TOKEN: ${{ secrets.BUILD_TOKEN }}
+        run: |
+          sudo apt-get install -y jq
+          REPO_API="${GITHUB_SERVER_URL}/api/v1/repos/${GITHUB_REPOSITORY}"
+
+          # Use version tag for tag pushes, "latest" for main
+          if [[ "${GITHUB_REF}" == refs/tags/* ]]; then
+            TAG="${GITHUB_REF#refs/tags/}"
+            RELEASE_NAME="Voice to Notes ${TAG}"
+            PRERELEASE=false
+          else
+            TAG="latest"
+            RELEASE_NAME="Voice to Notes (Latest Build)"
+            PRERELEASE=true
+          fi
+
+          echo "Release tag: ${TAG}"
+
+          RELEASE_ID=$(curl -s -H "Authorization: token ${BUILD_TOKEN}" \
+            "${REPO_API}/releases/tags/${TAG}" | jq -r '.id // empty')
+
+          if [ -z "${RELEASE_ID}" ]; then
+            RELEASE_ID=$(curl -s -X POST \
+              -H "Authorization: token ${BUILD_TOKEN}" \
+              -H "Content-Type: application/json" \
+              -d "{\"tag_name\": \"${TAG}\", \"name\": \"${RELEASE_NAME}\", \"body\": \"Automated build.\", \"draft\": false, \"prerelease\": ${PRERELEASE}}" \
+              "${REPO_API}/releases" | jq -r '.id')
+          fi
+
+          echo "Release ID: ${RELEASE_ID}"
+          if [ "${RELEASE_ID}" = "null" ] || [ -z "${RELEASE_ID}" ]; then
+            echo "ERROR: Failed to create/find release."
+            exit 1
+          fi
+
+          find src-tauri/target/release/bundle -type f -name "*.deb" | while IFS= read -r file; do
+            filename=$(basename "$file")
+            encoded_name=$(echo "$filename" | sed 's/ /%20/g')
+            echo "Uploading ${filename} ($(du -h "$file" | cut -f1))..."
+
+            ASSET_ID=$(curl -s -H "Authorization: token ${BUILD_TOKEN}" \
+              "${REPO_API}/releases/${RELEASE_ID}/assets" | jq -r ".[] | select(.name == \"${filename}\") | .id // empty")
+            if [ -n "${ASSET_ID}" ]; then
+              curl -s -X DELETE -H "Authorization: token ${BUILD_TOKEN}" \
+                "${REPO_API}/releases/${RELEASE_ID}/assets/${ASSET_ID}"
+            fi
+
+            HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" -X POST \
+              -H "Authorization: token ${BUILD_TOKEN}" \
+              -H "Content-Type: application/octet-stream" \
+              -T "$file" \
+              "${REPO_API}/releases/${RELEASE_ID}/assets?name=${encoded_name}")
+            echo "Upload response: HTTP ${HTTP_CODE}"
+          done

.gitea/workflows/build-macos.yml

@@ -0,0 +1,122 @@
+name: Build macOS
+
+on:
+  push:
+    branches: [main]
+    tags: ["v*"]
+  pull_request:
+    branches: [main]
+
+env:
+  PYTHON_VERSION: "3.11"
+  NODE_VERSION: "20"
+  TARGET: aarch64-apple-darwin
+
+jobs:
+  build:
+    name: Build (macOS)
+    runs-on: macos-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      # ── Python sidecar ──
+      - name: Install uv
+        run: |
+          if command -v uv &> /dev/null; then
+            echo "uv already installed: $(uv --version)"
+          else
+            curl -LsSf https://astral.sh/uv/install.sh | sh
+            echo "$HOME/.local/bin" >> $GITHUB_PATH
+          fi
+
+      - name: Install ffmpeg
+        run: brew install ffmpeg
+
+      - name: Set up Python
+        run: uv python install ${{ env.PYTHON_VERSION }}
+
+      - name: Build sidecar
+        working-directory: python
+        run: uv run --python ${{ env.PYTHON_VERSION }} python build_sidecar.py --cpu-only
+
+      - name: Package sidecar for Tauri
+        run: |
+          cd python/dist/voice-to-notes-sidecar && zip -r ../../../src-tauri/sidecar.zip .
+
+      # ── Tauri app ──
+      - name: Set up Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: ${{ env.NODE_VERSION }}
+
+      - name: Install Rust stable
+        run: |
+          curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --default-toolchain stable
+          echo "$HOME/.cargo/bin" >> $GITHUB_PATH
+
+      - name: Install system dependencies
+        run: brew install --quiet create-dmg || true
+
+      - name: Install npm dependencies
+        run: npm ci
+
+      - name: Build Tauri app
+        run: npm run tauri build
+
+      # ── Release ──
+      - name: Upload to release
+        if: github.ref == 'refs/heads/main' || startsWith(github.ref, 'refs/tags/v')
+        env:
+          BUILD_TOKEN: ${{ secrets.BUILD_TOKEN }}
+        run: |
+          REPO_API="${GITHUB_SERVER_URL}/api/v1/repos/${GITHUB_REPOSITORY}"
+
+          # Use version tag for tag pushes, "latest" for main
+          if [[ "${GITHUB_REF}" == refs/tags/* ]]; then
+            TAG="${GITHUB_REF#refs/tags/}"
+            RELEASE_NAME="Voice to Notes ${TAG}"
+            PRERELEASE=false
+          else
+            TAG="latest"
+            RELEASE_NAME="Voice to Notes (Latest Build)"
+            PRERELEASE=true
+          fi
+
+          echo "Release tag: ${TAG}"
+
+          RELEASE_ID=$(curl -s -H "Authorization: token ${BUILD_TOKEN}" \
+            "${REPO_API}/releases/tags/${TAG}" | jq -r '.id // empty')
+
+          if [ -z "${RELEASE_ID}" ]; then
+            RELEASE_ID=$(curl -s -X POST \
+              -H "Authorization: token ${BUILD_TOKEN}" \
+              -H "Content-Type: application/json" \
+              -d "{\"tag_name\": \"${TAG}\", \"name\": \"${RELEASE_NAME}\", \"body\": \"Automated build.\", \"draft\": false, \"prerelease\": ${PRERELEASE}}" \
+              "${REPO_API}/releases" | jq -r '.id')
+          fi
+
+          echo "Release ID: ${RELEASE_ID}"
+          if [ "${RELEASE_ID}" = "null" ] || [ -z "${RELEASE_ID}" ]; then
+            echo "ERROR: Failed to create/find release."
+            exit 1
+          fi
+
+          find src-tauri/target/release/bundle -type f -name "*.dmg" | while IFS= read -r file; do
+            filename=$(basename "$file")
+            encoded_name=$(echo "$filename" | sed 's/ /%20/g')
+            echo "Uploading ${filename} ($(du -h "$file" | cut -f1))..."
+
+            ASSET_ID=$(curl -s -H "Authorization: token ${BUILD_TOKEN}" \
+              "${REPO_API}/releases/${RELEASE_ID}/assets" | jq -r ".[] | select(.name == \"${filename}\") | .id // empty")
+            if [ -n "${ASSET_ID}" ]; then
+              curl -s -X DELETE -H "Authorization: token ${BUILD_TOKEN}" \
+                "${REPO_API}/releases/${RELEASE_ID}/assets/${ASSET_ID}"
+            fi
+
+            HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" -X POST \
+              -H "Authorization: token ${BUILD_TOKEN}" \
+              -H "Content-Type: application/octet-stream" \
+              -T "$file" \
+              "${REPO_API}/releases/${RELEASE_ID}/assets?name=${encoded_name}")
+            echo "Upload response: HTTP ${HTTP_CODE}"
+          done

.gitea/workflows/build-windows.yml

@@ -0,0 +1,144 @@
+name: Build Windows
+
+on:
+  push:
+    branches: [main]
+    tags: ["v*"]
+  pull_request:
+    branches: [main]
+
+env:
+  PYTHON_VERSION: "3.11"
+  NODE_VERSION: "20"
+  TARGET: x86_64-pc-windows-msvc
+
+jobs:
+  build:
+    name: Build (Windows)
+    runs-on: windows-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      # ── Python sidecar ──
+      - name: Install uv
+        shell: powershell
+        run: |
+          if (Get-Command uv -ErrorAction SilentlyContinue) {
+            Write-Host "uv already installed: $(uv --version)"
+          } else {
+            irm https://astral.sh/uv/install.ps1 | iex
+            echo "$env:USERPROFILE\.local\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          }
+
+      - name: Install ffmpeg
+        shell: powershell
+        run: choco install ffmpeg -y
+
+      - name: Set up Python
+        shell: powershell
+        run: uv python install ${{ env.PYTHON_VERSION }}
+
+      - name: Build sidecar
+        shell: powershell
+        working-directory: python
+        run: uv run --python ${{ env.PYTHON_VERSION }} python build_sidecar.py --cpu-only
+
+      - name: Package sidecar for Tauri
+        shell: powershell
+        run: |
+          Compress-Archive -Path python\dist\voice-to-notes-sidecar\* -DestinationPath src-tauri\sidecar.zip
+
+      # ── Tauri app ──
+      - name: Set up Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: ${{ env.NODE_VERSION }}
+
+      - name: Install Rust stable
+        shell: powershell
+        run: |
+          if (Get-Command rustup -ErrorAction SilentlyContinue) {
+            rustup default stable
+          } else {
+            Invoke-WebRequest -Uri https://win.rustup.rs/x86_64 -OutFile rustup-init.exe
+            .\rustup-init.exe -y --default-toolchain stable
+            echo "$env:USERPROFILE\.cargo\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          }
+
+      - name: Install npm dependencies
+        shell: powershell
+        run: npm ci
+
+      - name: Build Tauri app
+        shell: powershell
+        run: npm run tauri build
+
+      # ── Release ──
+      - name: Upload to release
+        if: github.ref == 'refs/heads/main' || startsWith(github.ref, 'refs/tags/v')
+        shell: powershell
+        env:
+          BUILD_TOKEN: ${{ secrets.BUILD_TOKEN }}
+        run: |
+          $REPO_API = "${{ github.server_url }}/api/v1/repos/${{ github.repository }}"
+          $Headers = @{ "Authorization" = "token $env:BUILD_TOKEN" }
+
+          # Use version tag for tag pushes, "latest" for main
+          $REF = "${{ github.ref }}"
+          if ($REF.StartsWith("refs/tags/")) {
+            $TAG = $REF.Replace("refs/tags/", "")
+            $RELEASE_NAME = "Voice to Notes ${TAG}"
+            $PRERELEASE = $false
+          } else {
+            $TAG = "latest"
+            $RELEASE_NAME = "Voice to Notes (Latest Build)"
+            $PRERELEASE = $true
+          }
+
+          Write-Host "Release tag: ${TAG}"
+
+          try {
+            $release = Invoke-RestMethod -Uri "${REPO_API}/releases/tags/${TAG}" -Headers $Headers -ErrorAction Stop
+            $RELEASE_ID = $release.id
+          } catch {
+            $body = @{
+              tag_name = $TAG
+              name = $RELEASE_NAME
+              body = "Automated build."
+              draft = $false
+              prerelease = $PRERELEASE
+            } | ConvertTo-Json
+            $release = Invoke-RestMethod -Uri "${REPO_API}/releases" -Method Post -Headers $Headers -ContentType "application/json" -Body $body
+            $RELEASE_ID = $release.id
+          }
+
+          Write-Host "Release ID: ${RELEASE_ID}"
+
+          Get-ChildItem -Path src-tauri\target\release\bundle -Recurse -Include *.msi,*-setup.exe | ForEach-Object {
+            $filename = $_.Name
+            $encodedName = [System.Uri]::EscapeDataString($filename)
+            $size = [math]::Round($_.Length / 1MB, 1)
+            Write-Host "Uploading ${filename} (${size} MB)..."
+
+            try {
+              $assets = Invoke-RestMethod -Uri "${REPO_API}/releases/${RELEASE_ID}/assets" -Headers $Headers
+              $existing = $assets | Where-Object { $_.name -eq $filename }
+              if ($existing) {
+                Invoke-RestMethod -Uri "${REPO_API}/releases/${RELEASE_ID}/assets/$($existing.id)" -Method Delete -Headers $Headers
+              }
+            } catch {}
+
+            # Use curl for streaming upload (Invoke-RestMethod fails on large files)
+            $uploadUrl = "${REPO_API}/releases/${RELEASE_ID}/assets?name=${encodedName}"
+            $result = curl.exe --fail --silent --show-error `
+              -X POST `
+              -H "Authorization: token $env:BUILD_TOKEN" `
+              -H "Content-Type: application/octet-stream" `
+              --data-binary "@$($_.FullName)" `
+              "$uploadUrl" 2>&1
+            if ($LASTEXITCODE -eq 0) {
+              Write-Host "Upload successful: ${filename}"
+            } else {
+              Write-Host "WARNING: Upload failed for ${filename}: ${result}"
+            }
+          }

.gitignore

@@ -46,3 +46,10 @@ Thumbs.db
 *.ogg
 *.flac
 !test/fixtures/*
+
+# Sidecar build artifacts
+src-tauri/binaries/*
+!src-tauri/binaries/.gitkeep
+src-tauri/sidecar.zip
+python/dist/
+python/build/

CLAUDE.md

@@ -8,7 +8,7 @@ Desktop app for transcribing audio/video with speaker identification. Runs local
 - **ML pipeline:** Python sidecar process (faster-whisper, pyannote.audio, wav2vec2)
 - **Database:** SQLite (via rusqlite in Rust)
 - **Local AI:** Bundled llama-server (llama.cpp) — default, no install needed
-- **Cloud AI providers:** LiteLLM, OpenAI, Anthropic (optional, user-configured)
+- **Cloud AI providers:** OpenAI, Anthropic, OpenAI-compatible endpoints (optional, user-configured)
 - **Caption export:** pysubs2 (Python)
 - **Audio UI:** wavesurfer.js
 - **Transcript editor:** TipTap (ProseMirror)
@@ -40,7 +40,13 @@ docs/                   # Architecture and design documents
 - Database: UUIDs as primary keys (TEXT type in SQLite)
 - All timestamps in milliseconds (integer) relative to media file start
 
+## Distribution
+- Python sidecar is frozen via PyInstaller into a standalone binary for distribution
+- Tauri bundles the sidecar via `externalBin` — no Python required for end users
+- CI/CD builds on Gitea Actions (Linux, Windows, macOS ARM)
+- Dev mode uses system Python (`VOICE_TO_NOTES_DEV=1` or debug builds)
+
 ## Platform Targets
-- Linux (primary development target)
-- Windows (must work, tested before release)
-- macOS (future, not yet targeted)
+- Linux x86_64 (primary development target)
+- Windows x86_64
+- macOS aarch64 (Apple Silicon)

README.md

@@ -2,28 +2,90 @@
 
 A desktop application that transcribes audio/video recordings with speaker identification, producing editable transcriptions with synchronized audio playback.
 
-## Goals
+## Features
 
-- **Speech-to-Text Transcription** — Accurately convert spoken audio from recordings into text
-- **Speaker Identification (Diarization)** — Detect and distinguish between different speakers in a conversation
-- **Speaker Naming** — Assign and persist speaker names/IDs across the transcription
-- **Synchronized Playback** — Click any transcribed text segment to play back the corresponding audio for review and correction
-- **Export Formats**
-  - Closed captioning files (SRT, VTT) for video
-  - Plain text documents with speaker labels
-- **AI Integration** — Connect to AI providers to ask questions about the conversation and generate condensed notes/summaries
+- **Speech-to-Text Transcription** — Accurate transcription via faster-whisper (Whisper models) with word-level timestamps
+- **Speaker Identification (Diarization)** — Detect and distinguish between speakers using pyannote.audio
+- **Synchronized Playback** — Click any word to seek to that point in the audio (Web Audio API for instant playback)
+- **AI Integration** — Ask questions about your transcript via OpenAI, Anthropic, or any OpenAI-compatible API (LiteLLM proxies, Ollama, vLLM)
+- **Export Formats** — SRT, WebVTT, ASS captions, plain text, and Markdown with speaker labels
+- **Cross-Platform** — Builds for Linux, Windows, and macOS (Apple Silicon)
 
 ## Platform Support
 
-| Platform | Status |
-|----------|--------|
-| Linux    | Planned (initial target) |
-| Windows  | Planned (initial target) |
-| macOS    | Future (pending hardware) |
+| Platform | Architecture | Status |
+|----------|-------------|--------|
+| Linux    | x86_64      | Supported |
+| Windows  | x86_64      | Supported |
+| macOS    | ARM (Apple Silicon) | Supported |
 
-## Project Status
+## Tech Stack
 
-**Early planning phase** — Architecture and technology decisions in progress.
+- **Desktop shell:** Tauri v2 (Rust backend + Svelte 5 / TypeScript frontend)
+- **ML pipeline:** Python sidecar (faster-whisper, pyannote.audio) — frozen via PyInstaller for distribution
+- **Audio playback:** wavesurfer.js with Web Audio API backend
+- **AI providers:** OpenAI, Anthropic, OpenAI-compatible endpoints (local or remote)
+- **Local AI:** Bundled llama-server (llama.cpp)
+- **Caption export:** pysubs2
+
+## Development
+
+### Prerequisites
+
+- Node.js 20+
+- Rust (stable)
+- Python 3.11+ with ML dependencies
+- System: `libgtk-3-dev`, `libwebkit2gtk-4.1-dev` (Linux)
+
+### Getting Started
+
+```bash
+# Install frontend dependencies
+npm install
+
+# Install Python sidecar dependencies
+cd python && pip install -e . && cd ..
+
+# Run in dev mode (uses system Python for the sidecar)
+npm run tauri:dev
+```
+
+### Building for Distribution
+
+```bash
+# Build the frozen Python sidecar
+npm run sidecar:build
+
+# Build the Tauri app (requires sidecar in src-tauri/binaries/)
+npm run tauri build
+```
+
+### CI/CD
+
+Gitea Actions workflows are in `.gitea/workflows/`. The build pipeline:
+
+1. **Build sidecar** — PyInstaller-frozen Python binary per platform (CPU-only PyTorch)
+2. **Build Tauri app** — Bundles the sidecar via `externalBin`, produces .deb/.AppImage (Linux), .msi (Windows), .dmg (macOS)
+
+#### Required Secrets
+
+| Secret | Purpose | Required? |
+|--------|---------|-----------|
+| `TAURI_SIGNING_PRIVATE_KEY` | Signs Tauri update bundles | Optional (for auto-updates) |
+
+No other secrets are needed for building. AI provider API keys and HuggingFace tokens are configured by end users in the app's Settings.
+
+### Project Structure
+
+```
+src/                    # Svelte 5 frontend
+src-tauri/              # Rust backend (Tauri commands, sidecar manager, SQLite)
+python/                 # Python sidecar (transcription, diarization, AI)
+  voice_to_notes/       # Python package
+  build_sidecar.py      # PyInstaller build script
+  voice_to_notes.spec   # PyInstaller spec
+.gitea/workflows/       # Gitea Actions CI/CD
+```
 
 ## License
 

RESEARCH_REPORT.md

@@ -0,0 +1,380 @@
+# Voice to Notes: Speech-to-Text and Speaker Diarization Research Report
+
+**Date:** 2026-02-26
+
+---
+
+## Table of Contents
+
+1. [Speech-to-Text Engines](#1-speech-to-text-engines)
+2. [Speaker Diarization](#2-speaker-diarization)
+3. [Combined Pipelines](#3-combined-pipelines)
+4. [Final Recommendations](#4-final-recommendations)
+
+---
+
+## 1. Speech-to-Text Engines
+
+### 1.1 OpenAI Whisper / whisper.cpp
+
+**Overview:** Whisper is OpenAI's general-purpose speech recognition model trained on 680,000 hours of multilingual data. whisper.cpp is a pure C/C++ port by Georgi Gerganov (ggml project) that removes the Python/PyTorch dependency entirely.
+
+| Criterion | Assessment |
+|---|---|
+| **Accuracy** | State-of-the-art. Large-v3 achieves ~2.7% WER on clean audio, ~7.9% on mixed real-world audio. Large-v3-turbo achieves comparable accuracy (~7.75% WER) at much faster speed by reducing decoder layers from 32 to 4. |
+| **Speed** | whisper.cpp with quantization (Q5_K_M) runs efficiently on CPU. GPU acceleration available via CUDA (NVIDIA), Vulkan (cross-vendor), Metal (Apple Silicon), and OpenVINO (Intel). Real-time or faster on modern hardware with medium/small models. |
+| **Language Support** | 99 languages. |
+| **Ease of Integration** | whisper.cpp: C/C++ library with C API, bindings available for many languages. No Python runtime needed. Straightforward to embed in a desktop app. |
+| **License** | MIT (both Whisper and whisper.cpp). |
+| **GPU Acceleration** | CUDA, Vulkan, Metal, OpenVINO, CoreML. Broad hardware coverage. |
+| **Word-Level Timestamps** | Supported, but derived from forced alignment on decoded text rather than internal attention weights. Can drift 300-800ms on complex utterances. Acceptable for many use cases but not forensic-grade. |
+
+**Verdict:** Best option for a native desktop app that needs to minimize dependencies. The C/C++ nature of whisper.cpp makes it ideal for embedding in an Electron, Qt, or Tauri application.
+
+---
+
+### 1.2 faster-whisper
+
+**Overview:** A Python reimplementation of Whisper using CTranslate2, a high-performance C++ inference engine for Transformer models. Up to 4x faster than stock Whisper with the same accuracy, and lower memory usage.
+
+| Criterion | Assessment |
+|---|---|
+| **Accuracy** | Identical to Whisper (same models, full fidelity). |
+| **Speed** | Up to 4x faster than stock Whisper. ~20x realtime with GPU. 8-bit quantization available on both CPU and GPU. |
+| **Language Support** | 99 languages (same Whisper models). |
+| **Ease of Integration** | Python library. Requires Python runtime. Excellent for Python-based or Python-embedded apps. Rich API with access to Whisper's tokenizer, alignment algorithms, and confidence scoring. |
+| **License** | MIT. |
+| **GPU Acceleration** | NVIDIA CUDA, AMD ROCm (via CTranslate2). CPU backends: Intel MKL, oneDNN, OpenBLAS, Ruy. |
+| **Word-Level Timestamps** | **Best-in-class among Whisper variants.** Native alignment from the model's internals plus optional wav2vec2 alignment for even better precision. |
+
+**Verdict:** Best choice if your app can embed a Python runtime (or run a Python sidecar process). Provides the most precise word-level timestamps of any Whisper variant, which is critical for synchronized playback. The trade-off is the Python dependency.
+
+---
+
+### 1.3 Vosk
+
+**Overview:** A lightweight, Kaldi-based offline speech recognition toolkit. Optimized for efficiency and small footprint.
+
+| Criterion | Assessment |
+|---|---|
+| **Accuracy** | Good but noticeably below Whisper-class models. Baseline WER can be ~20%+ depending on audio conditions, improvable to ~12% with domain-specific language model adaptation. |
+| **Speed** | Very fast, even on low-end hardware. Supports real-time streaming natively. |
+| **Language Support** | 20+ languages with pre-trained models. |
+| **Ease of Integration** | Excellent. APIs for Python, Java, C#, JavaScript, Node.js, and more. Models are ~50MB. |
+| **License** | Apache 2.0. |
+| **GPU Acceleration** | Not required (runs efficiently on CPU). No GPU acceleration. |
+| **Word-Level Timestamps** | Yes, provides word-level timestamps with start/end times and confidence in JSON output. |
+
+**Verdict:** Best for extremely resource-constrained scenarios or as a lightweight fallback. Not recommended as the primary engine for a quality-focused transcription app due to lower accuracy compared to Whisper-based solutions.
+
+---
+
+### 1.4 Coqui STT
+
+**Overview:** Fork of Mozilla DeepSpeech. The Coqui company shut down in early 2024. The code remains available as open source, but the project is no longer maintained and the Model Zoo is offline.
+
+| Criterion | Assessment |
+|---|---|
+| **Accuracy** | Below Whisper. Was competitive in the DeepSpeech era but has fallen behind. |
+| **Speed** | Moderate. |
+| **Language Support** | Limited compared to Whisper. |
+| **Ease of Integration** | Python and native bindings available, but stale dependencies. |
+| **License** | MPL 2.0. |
+| **GPU Acceleration** | TensorFlow-based GPU support. |
+| **Word-Level Timestamps** | Supported via metadata output. |
+
+**Verdict:** **Not recommended.** The project is discontinued. No active maintenance, no security patches, no model improvements. Use Whisper-based alternatives instead.
+
+---
+
+### 1.5 Other Notable Options
+
+#### Whisper Large-v3-turbo
+OpenAI's latest Whisper variant (October 2024). Reduces decoder layers from 32 to 4 while maintaining accuracy close to large-v3. Achieves 216x realtime speed. Available in both whisper.cpp and faster-whisper.
+
+#### NVIDIA NeMo ASR
+Production-grade ASR with Conformer-CTC and Conformer-Transducer models. Best accuracy in some benchmarks but heavy dependency on NVIDIA ecosystem. Apache 2.0 license. Overkill for a desktop app unless targeting NVIDIA GPU users specifically.
+
+#### Wav2Vec2 (Meta)
+Strong accuracy when fine-tuned for specific domains. Good for real-time streaming. Often used as an alignment model rather than primary STT. MIT license.
+
+---
+
+### STT Summary Comparison
+
+| Feature | whisper.cpp | faster-whisper | Vosk | Coqui STT |
+|---|---|---|---|---|
+| **Accuracy** | Excellent | Excellent | Good | Fair |
+| **Speed** | Fast | Very Fast | Very Fast | Moderate |
+| **Languages** | 99 | 99 | 20+ | Limited |
+| **Word Timestamps** | Yes (some drift) | Yes (precise) | Yes | Yes |
+| **GPU Support** | CUDA/Vulkan/Metal | CUDA/ROCm | CPU only | TensorFlow |
+| **License** | MIT | MIT | Apache 2.0 | MPL 2.0 |
+| **Dependencies** | None (C/C++) | Python + CTranslate2 | Minimal | Python + TF |
+| **Actively Maintained** | Yes | Yes | Yes | **No** |
+| **Desktop-Friendly** | Excellent | Good | Excellent | Poor |
+
+---
+
+## 2. Speaker Diarization
+
+### 2.1 pyannote.audio
+
+**Overview:** The leading open-source speaker diarization toolkit. Recently released version 4.0 with the "community-1" model, which significantly outperforms the previous 3.1 across all metrics.
+
+| Criterion | Assessment |
+|---|---|
+| **Accuracy** | Best-in-class open source. DER (Diarization Error Rate) ~11-19% on standard benchmarks. Community-1 model is a major leap over 3.1. |
+| **Pre-recorded Audio** | Full support. Designed for both offline and streaming use. |
+| **Ease of Integration** | Python library with PyTorch backend. Simple pipeline API: `pipeline("audio.wav")` returns speaker segments. Can run fully offline once models are downloaded. |
+| **Combinable with STT** | Yes. WhisperX and whisper-diarization both use pyannote as their diarization backend. Well-established integration patterns. |
+| **License** | Code: MIT. Models: speaker-diarization-3.1 is MIT; community-1 is CC-BY-4.0. Both allow commercial use. |
+| **GPU Support** | Yes, PyTorch CUDA. Can also run on CPU (slower but functional). |
+
+**Verdict:** Clear first choice for diarization. Most accurate, best maintained, largest community, and proven integration with Whisper-based STT. The community-1 model under CC-BY-4.0 is permissive enough for commercial desktop apps.
+
+---
+
+### 2.2 NVIDIA NeMo Speaker Diarization
+
+**Overview:** Part of NVIDIA's NeMo framework. Offers two approaches: end-to-end Sortformer Diarizer and cascaded pipeline (MarbleNet VAD + TitaNet embeddings + Multi-Scale Diarization Decoder).
+
+| Criterion | Assessment |
+|---|---|
+| **Accuracy** | Competitive with or slightly better than pyannote in some benchmarks. Sortformer is state-of-the-art. |
+| **Pre-recorded Audio** | Full support. Also has streaming Sortformer for real-time. |
+| **Ease of Integration** | Heavy. NeMo is a large framework with many dependencies. Requires NVIDIA GPU for practical use. Complex configuration via YAML files. |
+| **Combinable with STT** | Yes. NeMo includes its own ASR models. Can combine diarization with NeMo ASR in a single pipeline. |
+| **License** | Apache 2.0. |
+| **GPU Support** | NVIDIA GPU required for practical performance. |
+
+**Verdict:** Best accuracy in some scenarios, but the heavy NVIDIA dependency and complex setup make it poorly suited for a consumer desktop app that must work across hardware. Good option if you can offer it as an optional backend for users with NVIDIA GPUs.
+
+---
+
+### 2.3 SpeechBrain
+
+**Overview:** An open-source, all-in-one conversational AI toolkit built on PyTorch. Covers ASR, speaker identification, diarization, speech enhancement, and more.
+
+| Criterion | Assessment |
+|---|---|
+| **Accuracy** | Good, though generally slightly behind pyannote on diarization-specific benchmarks. |
+| **Pre-recorded Audio** | Full support. |
+| **Ease of Integration** | Moderate. PyTorch-based. Well-documented but the "kitchen sink" approach means you pull in a large framework even if you only need diarization. |
+| **Combinable with STT** | Yes. Has its own ASR components. Can build end-to-end pipelines within the framework. |
+| **License** | Apache 2.0. |
+| **GPU Support** | PyTorch CUDA. |
+
+**Verdict:** Good option if you want a single framework for everything (ASR + diarization + enhancement). However, for diarization specifically, pyannote is more focused and generally more accurate. SpeechBrain is better suited for teams that want deep customization of the diarization pipeline.
+
+---
+
+### 2.4 Resemblyzer
+
+**Overview:** A Python library by Resemble AI for extracting speaker embeddings using a GE2E (Generalized End-to-End) model. Primarily a speaker verification/comparison tool, not a full diarization system.
+
+| Criterion | Assessment |
+|---|---|
+| **Accuracy** | Moderate. The underlying model is older and less accurate than pyannote or NeMo embeddings. |
+| **Pre-recorded Audio** | Yes, but you must build your own clustering/segmentation logic on top. |
+| **Ease of Integration** | Simple API for embedding extraction. But no built-in diarization pipeline; you need to implement VAD, segmentation, and clustering yourself. |
+| **Combinable with STT** | Manually, with significant custom code. |
+| **License** | Apache 2.0. |
+| **GPU Support** | PyTorch (optional). |
+| **Maintenance Status** | **Inactive.** No new releases or meaningful updates in over 12 months. |
+
+**Verdict:** **Not recommended** for new projects. It is essentially unmaintained and provides only embeddings, not a complete diarization solution. pyannote provides better embeddings and a complete pipeline.
+
+---
+
+### Diarization Summary Comparison
+
+| Feature | pyannote.audio | NeMo | SpeechBrain | Resemblyzer |
+|---|---|---|---|---|
+| **Accuracy (DER)** | ~11-19% | ~10-18% | ~13-20% | N/A (not a full system) |
+| **Complete Pipeline** | Yes | Yes | Yes | No (embeddings only) |
+| **Ease of Setup** | Easy | Complex | Moderate | Easy (but incomplete) |
+| **License** | MIT / CC-BY-4.0 | Apache 2.0 | Apache 2.0 | Apache 2.0 |
+| **GPU Required** | No (recommended) | Practically yes | No (recommended) | No |
+| **Actively Maintained** | Yes (v4.0, Feb 2026) | Yes | Yes | No |
+| **Desktop-Friendly** | Good | Poor | Moderate | N/A |
+
+---
+
+## 3. Combined Pipelines (STT + Diarization)
+
+### 3.1 WhisperX
+
+**Overview:** The most mature combined pipeline. Integrates faster-whisper (STT) + wav2vec2 (alignment) + pyannote.audio (diarization) into a single workflow.
+
+**How it works:**
+1. **Transcription:** faster-whisper transcribes audio into coarse utterance-level segments with batched inference (~70x realtime with large-v2).
+2. **Forced Alignment:** wav2vec2 refines timestamps to precise word-level accuracy.
+3. **Diarization:** pyannote.audio segments the audio by speaker.
+4. **Alignment:** Word-level timestamps from step 2 are aligned with speaker segments from step 3, assigning each word to a speaker.
+
+**Strengths:**
+- Best word-level timestamp accuracy of any open-source solution.
+- Speaker labels mapped to individual words.
+- Handles long audio files through intelligent chunking.
+- Active development, large community.
+
+**Weaknesses:**
+- Python-only. Requires Python runtime with PyTorch, faster-whisper, and pyannote dependencies.
+- Significant memory usage (multiple models loaded simultaneously).
+- Pyannote model download requires accepting license on Hugging Face (one-time).
+
+**License:** BSD-4-Clause (WhisperX itself); dependencies are MIT/Apache.
+
+---
+
+### 3.2 whisper-diarization (by MahmoudAshraf97)
+
+**Overview:** An alternative combined pipeline using Whisper + pyannote for diarization. Simpler than WhisperX but with fewer features.
+
+**Strengths:**
+- Straightforward Python script approach.
+- Uses pyannote for diarization.
+- Easier to understand and modify.
+
+**Weaknesses:**
+- Less optimized than WhisperX.
+- Fewer alignment options.
+
+---
+
+### 3.3 NVIDIA NeMo End-to-End
+
+**Overview:** NeMo can run ASR and diarization in a single framework. The Sortformer model handles diarization end-to-end, and NeMo ASR handles transcription.
+
+**Strengths:**
+- Single framework, no glue code between separate libraries.
+- State-of-the-art accuracy.
+- Streaming support with Streaming Sortformer.
+
+**Weaknesses:**
+- Requires NVIDIA GPU.
+- Heavy framework, not consumer-desktop friendly.
+- Complex configuration.
+
+---
+
+### 3.4 Aligning Diarization with Transcription Timestamps
+
+The fundamental challenge: STT produces words with timestamps, while diarization produces speaker segments with timestamps. These must be merged.
+
+**Best Approach (used by WhisperX):**
+
+```
+1. Run STT -> get words with [start_time, end_time] per word
+2. Run diarization -> get speaker segments [speaker_id, start_time, end_time]
+3. For each word, find which speaker segment it falls into:
+   - Use the word's midpoint timestamp
+   - Assign the word to whichever speaker segment contains that midpoint
+   - Handle edge cases (words spanning segment boundaries) with majority overlap
+```
+
+**Alignment quality depends on:**
+- **Word timestamp precision:** faster-whisper with wav2vec2 alignment provides the best precision. whisper.cpp timestamps can drift 300-800ms, which can cause mis-attribution at speaker boundaries.
+- **Diarization segment precision:** pyannote.audio community-1 provides the tightest speaker boundaries.
+- **Overlap handling:** In conversations where speakers overlap, both timestamps and diarization become less reliable. pyannote.audio 4.0 has specific overlapped speech detection.
+
+---
+
+## 4. Final Recommendations
+
+### Primary Recommendation: Two-Tier Architecture
+
+Given the "Voice to Notes" requirements (local-first, consumer hardware, word-level timestamps for synchronized playback, speaker identification), I recommend a **two-tier architecture**:
+
+#### Tier 1: Core Transcription Engine (C/C++)
+
+**Use whisper.cpp** as the primary STT engine.
+
+- No Python dependency for the core app.
+- Runs on all hardware (CPU, NVIDIA GPU, AMD GPU via Vulkan, Intel via OpenVINO).
+- MIT license with no restrictions.
+- Embed directly into your desktop app (Tauri, Qt, Electron with native addon).
+- Use the `large-v3-turbo` model as the default (best speed/accuracy trade-off for consumer hardware).
+- Offer `medium` and `small` models for lower-end hardware.
+- Word-level timestamps are adequate for playback synchronization (300-800ms drift is acceptable when the UI highlights the current phrase rather than individual words).
+
+#### Tier 2: Enhanced Pipeline (Python Sidecar)
+
+**Use faster-whisper + pyannote.audio** via a Python sidecar process for users who want speaker diarization and precise word-level alignment.
+
+- Ship a bundled Python environment (e.g., via PyInstaller or conda-pack).
+- Run the WhisperX-style pipeline: faster-whisper -> wav2vec2 alignment -> pyannote diarization.
+- Communicate with the main app via IPC (stdin/stdout JSON, local socket, or gRPC).
+- This gives the best word-level timestamps and speaker identification.
+- Optional: only install/download when user enables "Speaker Identification" feature.
+
+#### Model Selection Guide
+
+| User's Hardware | STT Model | Diarization |
+|---|---|---|
+| No GPU, 8GB RAM | whisper.cpp `small` (Q5_K_M) | pyannote on CPU (slower but works) |
+| No GPU, 16GB RAM | whisper.cpp `medium` (Q5_K_M) | pyannote on CPU |
+| NVIDIA GPU, 8GB+ VRAM | faster-whisper `large-v3-turbo` (int8) | pyannote on GPU |
+| NVIDIA GPU, 4GB VRAM | faster-whisper `medium` (int8) | pyannote on GPU |
+| Any hardware, speed priority | whisper.cpp `small` or `base` | Skip diarization |
+
+#### Optional Cloud Fallback
+
+For users who prefer cloud processing, integrate an optional cloud STT API (OpenAI Whisper API, AssemblyAI, or Deepgram) as a premium feature. This requires minimal code since the output format (words + timestamps + speakers) is the same regardless of backend.
+
+### Why Not Other Options?
+
+| Option | Reason to Skip |
+|---|---|
+| **Vosk** | Accuracy gap too large vs. Whisper. Only consider as a real-time streaming preview (show rough text while recording, then refine with Whisper afterward). |
+| **Coqui STT** | Discontinued. No future. |
+| **Resemblyzer** | Unmaintained, incomplete (no pipeline). |
+| **NeMo (full)** | Too heavy for consumer desktop. NVIDIA-only for practical use. |
+| **SpeechBrain** | Less accurate diarization than pyannote. Larger framework for less benefit. |
+
+### Recommended Technology Stack Summary
+
+```
+Desktop App Shell: Tauri (Rust) or Electron
+                         |
+        +----------------+----------------+
+        |                                 |
+  Core STT Engine                  Enhanced Pipeline
+  (whisper.cpp, C/C++)            (Python sidecar)
+        |                                 |
+  - Transcription              - faster-whisper (STT)
+  - Basic word timestamps      - wav2vec2 (alignment)
+  - No speaker ID              - pyannote.audio (diarization)
+                               - Precise word timestamps
+                               - Speaker identification
+```
+
+### Key Files and Repositories
+
+- **whisper.cpp:** https://github.com/ggml-org/whisper.cpp
+- **faster-whisper:** https://github.com/SYSTRAN/faster-whisper
+- **pyannote.audio:** https://github.com/pyannote/pyannote-audio
+- **WhisperX:** https://github.com/m-bain/whisperX
+- **whisper-diarization:** https://github.com/MahmoudAshraf97/whisper-diarization
+
+---
+
+## Sources
+
+- [OpenAI Whisper vs Vosk Comparison (Jamy AI)](https://www.jamy.ai/blog/openai-whisper-vs-other-open-source-transcription-models/)
+- [Top Open Source Transcription Software 2025 (Amical)](https://amical.ai/blog/open-source-transcription-software)
+- [Choosing Between Whisper Variants (Modal)](https://modal.com/blog/choosing-whisper-variants)
+- [whisper.cpp vs faster-whisper Practical Guide](https://www.alibaba.com/product-insights/a-practical-guide-to-choosing-between-whisper-cpp-and-faster-whisper-for-offline-transcription.html)
+- [Top 8 Open Source STT Options (AssemblyAI)](https://www.assemblyai.com/blog/top-open-source-stt-options-for-voice-applications)
+- [Best Speaker Diarization Models Compared 2026 (Brass Transcripts)](https://brasstranscripts.com/blog/speaker-diarization-models-comparison)
+- [Pyannote vs NeMo Comparison (La Javaness)](https://lajavaness.medium.com/comparing-state-of-the-art-speaker-diarization-frameworks-pyannote-vs-nemo-31a191c6300)
+- [Top Speaker Diarization Libraries 2026 (AssemblyAI)](https://www.assemblyai.com/blog/top-speaker-diarization-libraries-and-apis)
+- [pyannote.audio Community-1 Announcement](https://www.pyannote.ai/blog/community-1)
+- [pyannote/speaker-diarization-3.1 (Hugging Face)](https://huggingface.co/pyannote/speaker-diarization-3.1)
+- [Whisper Large-v3-turbo (Hugging Face)](https://huggingface.co/openai/whisper-large-v3-turbo)
+- [Best Open Source STT 2026 with Benchmarks (Northflank)](https://northflank.com/blog/best-open-source-speech-to-text-stt-model-in-2026-benchmarks)
+- [NVIDIA NeMo Speaker Diarization Docs](https://docs.nvidia.com/nemo-framework/user-guide/latest/nemotoolkit/asr/speaker_diarization/intro.html)
+- [faster-whisper (GitHub)](https://github.com/SYSTRAN/faster-whisper)
+- [WhisperX (GitHub)](https://github.com/m-bain/whisperX)
+- [Vosk Accuracy Guide](https://alphacephei.com/vosk/accuracy)

package.json

@@ -1,6 +1,6 @@
 {
   "name": "voice-to-notes",
-  "version": "0.1.0",
+  "version": "0.2.0",
   "description": "Desktop app for transcribing audio/video with speaker identification",
   "type": "module",
   "scripts": {
@@ -11,7 +11,9 @@
     "check:watch": "svelte-kit sync && svelte-check --tsconfig ./tsconfig.json --watch",
     "lint": "eslint .",
     "test": "vitest",
-    "tauri": "tauri"
+    "tauri": "tauri",
+    "tauri:dev": "VOICE_TO_NOTES_DEV=1 tauri dev",
+    "sidecar:build": "cd python && python3 build_sidecar.py"
   },
   "license": "MIT",
   "dependencies": {

python/build_sidecar.py

@@ -0,0 +1,245 @@
+#!/usr/bin/env python3
+"""Build the Voice to Notes sidecar as a standalone binary using PyInstaller.
+
+Usage:
+    python build_sidecar.py [--cpu-only]
+
+Produces a directory `dist/voice-to-notes-sidecar/` containing the frozen
+sidecar binary and all dependencies. The main binary is renamed to include
+the Tauri target triple for externalBin resolution.
+"""
+
+from __future__ import annotations
+
+import argparse
+import os
+import platform
+import shutil
+import stat
+import subprocess
+import sys
+import urllib.request
+import zipfile
+from pathlib import Path
+
+SCRIPT_DIR = Path(__file__).resolve().parent
+DIST_DIR = SCRIPT_DIR / "dist"
+BUILD_DIR = SCRIPT_DIR / "build"
+SPEC_FILE = SCRIPT_DIR / "voice_to_notes.spec"
+
+# Static ffmpeg download URLs (GPL-licensed builds)
+FFMPEG_URLS: dict[str, str] = {
+    "linux-x86_64": "https://johnvansickle.com/ffmpeg/releases/ffmpeg-release-amd64-static.tar.xz",
+    "darwin-x86_64": "https://evermeet.cx/ffmpeg/getrelease/zip",
+    "darwin-arm64": "https://evermeet.cx/ffmpeg/getrelease/zip",
+    "win32-x86_64": "https://www.gyan.dev/ffmpeg/builds/ffmpeg-release-essentials.zip",
+}
+
+
+def get_target_triple() -> str:
+    """Determine the Tauri-compatible target triple for the current platform."""
+    machine = platform.machine().lower()
+    system = platform.system().lower()
+
+    arch_map = {
+        "x86_64": "x86_64",
+        "amd64": "x86_64",
+        "aarch64": "aarch64",
+        "arm64": "aarch64",
+    }
+    arch = arch_map.get(machine, machine)
+
+    if system == "linux":
+        return f"{arch}-unknown-linux-gnu"
+    elif system == "darwin":
+        return f"{arch}-apple-darwin"
+    elif system == "windows":
+        return f"{arch}-pc-windows-msvc"
+    else:
+        return f"{arch}-unknown-{system}"
+
+
+def _has_uv() -> bool:
+    """Check if uv is available."""
+    try:
+        subprocess.run(["uv", "--version"], capture_output=True, check=True)
+        return True
+    except (FileNotFoundError, subprocess.CalledProcessError):
+        return False
+
+
+def create_venv_and_install(cpu_only: bool) -> Path:
+    """Create a fresh venv and install dependencies.
+
+    Uses uv if available (much faster), falls back to standard venv + pip.
+    """
+    venv_dir = BUILD_DIR / "sidecar-venv"
+    if venv_dir.exists():
+        shutil.rmtree(venv_dir)
+
+    use_uv = _has_uv()
+
+    if use_uv:
+        print(f"[build] Creating venv with uv at {venv_dir}")
+        subprocess.run(
+            ["uv", "venv", "--python", f"{sys.version_info.major}.{sys.version_info.minor}",
+             str(venv_dir)],
+            check=True,
+        )
+    else:
+        print(f"[build] Creating venv at {venv_dir}")
+        subprocess.run([sys.executable, "-m", "venv", str(venv_dir)], check=True)
+
+    # Determine python path inside venv
+    if sys.platform == "win32":
+        python = str(venv_dir / "Scripts" / "python.exe")
+    else:
+        python = str(venv_dir / "bin" / "python")
+
+    def pip_install(*args: str) -> None:
+        """Install packages. Pass package names and flags only, not 'install'."""
+        if use_uv:
+            # Use --python with the venv directory (not the python binary) for uv
+            subprocess.run(
+                ["uv", "pip", "install", "--python", str(venv_dir), *args],
+                check=True,
+            )
+        else:
+            subprocess.run([python, "-m", "pip", "install", *args], check=True)
+
+    if not use_uv:
+        # Upgrade pip (uv doesn't need this)
+        pip_install("--upgrade", "pip", "setuptools", "wheel")
+
+    # Install torch (CPU-only to avoid bundling ~2GB of CUDA libs)
+    if cpu_only:
+        print("[build] Installing PyTorch (CPU-only)")
+        pip_install(
+            "torch", "torchaudio",
+            "--index-url", "https://download.pytorch.org/whl/cpu",
+        )
+    else:
+        print("[build] Installing PyTorch (default, may include CUDA)")
+        pip_install("torch", "torchaudio")
+
+    # Install project and dev deps (includes pyinstaller)
+    print("[build] Installing project dependencies")
+    pip_install("-e", f"{SCRIPT_DIR}[dev]")
+
+    return Path(python)
+
+
+def run_pyinstaller(python: Path) -> Path:
+    """Run PyInstaller using the spec file."""
+    print("[build] Running PyInstaller")
+    subprocess.run(
+        [str(python), "-m", "PyInstaller", "--clean", "--noconfirm", str(SPEC_FILE)],
+        cwd=str(SCRIPT_DIR),
+        check=True,
+    )
+    output_dir = DIST_DIR / "voice-to-notes-sidecar"
+    if not output_dir.exists():
+        raise RuntimeError(f"PyInstaller output not found at {output_dir}")
+    return output_dir
+
+
+def download_ffmpeg(output_dir: Path) -> None:
+    """Download a static ffmpeg/ffprobe binary for the current platform."""
+    system = sys.platform
+    machine = platform.machine().lower()
+    if machine in ("amd64", "x86_64"):
+        machine = "x86_64"
+    elif machine in ("aarch64", "arm64"):
+        machine = "arm64"
+
+    key = f"{system}-{machine}"
+    if system == "win32":
+        key = f"win32-{machine}"
+    elif system == "linux":
+        key = f"linux-{machine}"
+
+    url = FFMPEG_URLS.get(key)
+    if not url:
+        print(f"[build] Warning: No ffmpeg download URL for platform {key}, skipping")
+        return
+
+    print(f"[build] Downloading ffmpeg for {key}")
+    tmp_path = output_dir / "ffmpeg_download"
+    try:
+        urllib.request.urlretrieve(url, str(tmp_path))
+
+        if url.endswith(".tar.xz"):
+            # Linux static build
+            import tarfile
+            with tarfile.open(str(tmp_path), "r:xz") as tar:
+                for member in tar.getmembers():
+                    basename = os.path.basename(member.name)
+                    if basename in ("ffmpeg", "ffprobe"):
+                        member.name = basename
+                        tar.extract(member, path=str(output_dir))
+                        dest = output_dir / basename
+                        dest.chmod(dest.stat().st_mode | stat.S_IEXEC)
+        elif url.endswith(".zip"):
+            with zipfile.ZipFile(str(tmp_path), "r") as zf:
+                for name in zf.namelist():
+                    basename = os.path.basename(name)
+                    if basename in ("ffmpeg", "ffprobe", "ffmpeg.exe", "ffprobe.exe"):
+                        data = zf.read(name)
+                        dest = output_dir / basename
+                        dest.write_bytes(data)
+                        if sys.platform != "win32":
+                            dest.chmod(dest.stat().st_mode | stat.S_IEXEC)
+        print("[build] ffmpeg downloaded successfully")
+    except Exception as e:
+        print(f"[build] Warning: Failed to download ffmpeg: {e}")
+    finally:
+        if tmp_path.exists():
+            tmp_path.unlink()
+
+
+def rename_binary(output_dir: Path, target_triple: str) -> None:
+    """Rename the main binary to include the target triple for Tauri."""
+    if sys.platform == "win32":
+        src = output_dir / "voice-to-notes-sidecar.exe"
+        dst = output_dir / f"voice-to-notes-sidecar-{target_triple}.exe"
+    else:
+        src = output_dir / "voice-to-notes-sidecar"
+        dst = output_dir / f"voice-to-notes-sidecar-{target_triple}"
+
+    if src.exists():
+        print(f"[build] Renaming {src.name} -> {dst.name}")
+        src.rename(dst)
+    else:
+        print(f"[build] Warning: Expected binary not found at {src}")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Build the Voice to Notes sidecar binary")
+    parser.add_argument(
+        "--cpu-only",
+        action="store_true",
+        default=True,
+        help="Install CPU-only PyTorch (default: True, avoids bundling CUDA)",
+    )
+    parser.add_argument(
+        "--with-cuda",
+        action="store_true",
+        help="Install PyTorch with CUDA support",
+    )
+    args = parser.parse_args()
+    cpu_only = not args.with_cuda
+
+    target_triple = get_target_triple()
+    print(f"[build] Target triple: {target_triple}")
+    print(f"[build] CPU-only: {cpu_only}")
+
+    python = create_venv_and_install(cpu_only)
+    output_dir = run_pyinstaller(python)
+    download_ffmpeg(output_dir)
+
+    print(f"\n[build] Done! Sidecar built at: {output_dir}")
+    print(f"[build] Copy directory to src-tauri/sidecar/ for Tauri resource bundling")
+
+
+if __name__ == "__main__":
+    main()

python/pyproject.toml

@@ -13,6 +13,8 @@ dependencies = [
     "faster-whisper>=1.1.0",
     "pyannote.audio>=3.1.0",
     "pysubs2>=1.7.0",
+    "openai>=1.0.0",
+    "anthropic>=0.20.0",
 ]
 
 [project.optional-dependencies]
@@ -20,6 +22,7 @@ dev = [
     "ruff>=0.8.0",
     "pytest>=8.0.0",
     "pytest-asyncio>=0.24.0",
+    "pyinstaller>=6.0",
 ]
 
 [tool.ruff]

python/tests/test_diarize.py

@@ -1,7 +1,13 @@
 """Tests for diarization service data structures and payload conversion."""
 
+import time
+from unittest.mock import MagicMock, patch
+
+import pytest
+
 from voice_to_notes.services.diarize import (
     DiarizationResult,
+    DiarizeService,
     SpeakerSegment,
     diarization_to_payload,
 )
@@ -31,3 +37,74 @@ def test_diarization_to_payload_empty():
     assert payload["num_speakers"] == 0
     assert payload["speaker_segments"] == []
     assert payload["speakers"] == []
+
+
+def test_diarize_threading_progress(monkeypatch):
+    """Test that diarization emits progress while running in background thread."""
+    # Track written messages
+    written_messages = []
+    def mock_write(msg):
+        written_messages.append(msg)
+
+    # Mock pipeline that takes ~5 seconds
+    def slow_pipeline(file_path, **kwargs):
+        time.sleep(5)
+        # Return a mock diarization result (use spec=object to prevent
+        # hasattr returning True for speaker_diarization)
+        mock_result = MagicMock(spec=[])
+        mock_track = MagicMock()
+        mock_track.start = 0.0
+        mock_track.end = 5.0
+        mock_result.itertracks = MagicMock(return_value=[(mock_track, None, "SPEAKER_00")])
+        return mock_result
+
+    mock_pipeline_obj = MagicMock()
+    mock_pipeline_obj.side_effect = slow_pipeline
+
+    service = DiarizeService()
+    service._pipeline = mock_pipeline_obj
+
+    with patch("voice_to_notes.services.diarize.write_message", mock_write):
+        result = service.diarize(
+            request_id="req-1",
+            file_path="/fake/audio.wav",
+            audio_duration_sec=60.0,
+        )
+
+    # Filter for diarizing progress messages (not loading_diarization or done)
+    diarizing_msgs = [
+        m for m in written_messages
+        if m.type == "progress" and m.payload.get("stage") == "diarizing"
+        and "elapsed" in m.payload.get("message", "")
+    ]
+
+    # Should have at least 1 progress message (5s sleep / 2s interval = ~2 messages)
+    assert len(diarizing_msgs) >= 1, (
+        f"Expected at least 1 diarizing progress message, got {len(diarizing_msgs)}"
+    )
+
+    # Progress percent should be between 20 and 85
+    for msg in diarizing_msgs:
+        pct = msg.payload["percent"]
+        assert 20 <= pct <= 85, f"Progress {pct} out of expected range 20-85"
+
+    # Result should be valid
+    assert result.num_speakers == 1
+    assert result.speakers == ["SPEAKER_00"]
+
+
+def test_diarize_threading_error_propagation(monkeypatch):
+    """Test that errors from the background thread are properly raised."""
+    mock_pipeline_obj = MagicMock()
+    mock_pipeline_obj.side_effect = RuntimeError("Pipeline crashed")
+
+    service = DiarizeService()
+    service._pipeline = mock_pipeline_obj
+
+    with patch("voice_to_notes.services.diarize.write_message", lambda m: None):
+        with pytest.raises(RuntimeError, match="Pipeline crashed"):
+            service.diarize(
+                request_id="req-1",
+                file_path="/fake/audio.wav",
+                audio_duration_sec=30.0,
+            )

python/tests/test_messages.py

@@ -3,8 +3,10 @@
 from voice_to_notes.ipc.messages import (
     IPCMessage,
     error_message,
+    partial_segment_message,
     progress_message,
     ready_message,
+    speaker_update_message,
 )
 
 
@@ -48,3 +50,16 @@ def test_ready_message():
     assert msg.type == "ready"
     assert msg.id == "system"
     assert "version" in msg.payload
+
+
+def test_partial_segment_message():
+    msg = partial_segment_message("req-1", {"index": 0, "text": "hello"})
+    assert msg.type == "pipeline.segment"
+    assert msg.payload["index"] == 0
+    assert msg.payload["text"] == "hello"
+
+
+def test_speaker_update_message():
+    msg = speaker_update_message("req-1", [{"index": 0, "speaker": "SPEAKER_00"}])
+    assert msg.type == "pipeline.speaker_update"
+    assert msg.payload["updates"][0]["speaker"] == "SPEAKER_00"

python/tests/test_pipeline.py

@@ -88,3 +88,18 @@ def test_merge_results_no_speaker_segments():
 
     result = service._merge_results(transcription, [])
     assert result.segments[0].speaker is None
+
+
+def test_speaker_update_generation():
+    """Test that speaker updates are generated after merge."""
+    result = PipelineResult(
+        segments=[
+            PipelineSegment(text="Hello", start_ms=0, end_ms=1000, speaker="SPEAKER_00"),
+            PipelineSegment(text="World", start_ms=1000, end_ms=2000, speaker="SPEAKER_01"),
+            PipelineSegment(text="Foo", start_ms=2000, end_ms=3000, speaker=None),
+        ],
+    )
+    updates = [{"index": i, "speaker": seg.speaker} for i, seg in enumerate(result.segments) if seg.speaker]
+    assert len(updates) == 2
+    assert updates[0] == {"index": 0, "speaker": "SPEAKER_00"}
+    assert updates[1] == {"index": 1, "speaker": "SPEAKER_01"}

python/tests/test_protocol.py

@@ -5,16 +5,23 @@ import json
 
 from voice_to_notes.ipc.messages import IPCMessage
 from voice_to_notes.ipc.protocol import read_message, write_message
+import voice_to_notes.ipc.protocol as protocol
 
 
-def test_write_message(capsys):
-    msg = IPCMessage(id="req-1", type="pong", payload={"ok": True})
-    write_message(msg)
-    captured = capsys.readouterr()
-    parsed = json.loads(captured.out.strip())
-    assert parsed["id"] == "req-1"
-    assert parsed["type"] == "pong"
-    assert parsed["payload"]["ok"] is True
+def test_write_message():
+    buf = io.StringIO()
+    # Temporarily replace the IPC output stream
+    old_out = protocol._ipc_out
+    protocol._ipc_out = buf
+    try:
+        msg = IPCMessage(id="req-1", type="pong", payload={"ok": True})
+        write_message(msg)
+        parsed = json.loads(buf.getvalue().strip())
+        assert parsed["id"] == "req-1"
+        assert parsed["type"] == "pong"
+        assert parsed["payload"]["ok"] is True
+    finally:
+        protocol._ipc_out = old_out
 
 
 def test_read_message(monkeypatch):

python/tests/test_transcribe.py

@@ -1,7 +1,10 @@
 """Tests for transcription service."""
 
+import inspect
+
 from voice_to_notes.services.transcribe import (
     SegmentResult,
+    TranscribeService,
     TranscriptionResult,
     WordResult,
     result_to_payload,
@@ -49,3 +52,149 @@ def test_result_to_payload_empty():
     assert payload["segments"] == []
     assert payload["language"] == ""
     assert payload["duration_ms"] == 0
+
+
+def test_on_segment_callback():
+    """Test that on_segment callback is invoked with correct SegmentResult and index."""
+    callback_args = []
+
+    def mock_callback(seg: SegmentResult, index: int):
+        callback_args.append((seg.text, index))
+
+    # Test that passing on_segment doesn't break the function signature
+    # (Full integration test would require mocking WhisperModel)
+    service = TranscribeService()
+    # Verify the parameter exists by checking the signature
+    sig = inspect.signature(service.transcribe)
+    assert "on_segment" in sig.parameters
+
+
+def test_progress_every_segment(monkeypatch):
+    """Verify a progress message is sent for every segment, not just every 5th."""
+    from unittest.mock import MagicMock, patch
+    from voice_to_notes.services.transcribe import TranscribeService
+
+    # Mock WhisperModel
+    mock_model = MagicMock()
+
+    # Create mock segments (8 of them to test > 5)
+    mock_segments = []
+    for i in range(8):
+        seg = MagicMock()
+        seg.start = i * 1.0
+        seg.end = (i + 1) * 1.0
+        seg.text = f"Segment {i}"
+        seg.words = []
+        mock_segments.append(seg)
+
+    # Mock info object
+    mock_info = MagicMock()
+    mock_info.language = "en"
+    mock_info.language_probability = 0.99
+    mock_info.duration = 8.0
+
+    mock_model.transcribe.return_value = (iter(mock_segments), mock_info)
+
+    # Track write_message calls
+    written_messages = []
+
+    def mock_write(msg):
+        written_messages.append(msg)
+
+    service = TranscribeService()
+    service._model = mock_model
+    service._current_model_name = "base"
+    service._current_device = "cpu"
+    service._current_compute_type = "int8"
+
+    with patch("voice_to_notes.services.transcribe.write_message", mock_write):
+        service.transcribe("req-1", "/fake/audio.wav")
+
+    # Filter for "transcribing" stage progress messages
+    transcribing_msgs = [
+        m for m in written_messages
+        if m.type == "progress" and m.payload.get("stage") == "transcribing"
+    ]
+
+    # Should have one per segment (8) + the initial "Starting transcription..." message
+    # The initial "Starting transcription..." is also stage "transcribing" — so 8 + 1 = 9
+    assert len(transcribing_msgs) >= 8, (
+        f"Expected at least 8 transcribing progress messages (one per segment), got {len(transcribing_msgs)}"
+    )
+
+
+def test_chunk_report_size_progress():
+    """Test CHUNK_REPORT_SIZE progress emission."""
+    from voice_to_notes.services.transcribe import CHUNK_REPORT_SIZE
+    assert CHUNK_REPORT_SIZE == 10
+
+
+def test_transcribe_chunked_with_mocked_ffmpeg(monkeypatch):
+    """Test transcribe_chunked with mocked ffmpeg/ffprobe and mocked WhisperModel."""
+    from unittest.mock import MagicMock, patch
+    from voice_to_notes.services.transcribe import TranscribeService, SegmentResult, WordResult
+
+    # Mock subprocess.run for ffprobe (returns duration of 700s = ~2 chunks at 300s each)
+    original_run = __import__("subprocess").run
+
+    def mock_subprocess_run(cmd, **kwargs):
+        if "ffprobe" in cmd:
+            result = MagicMock()
+            result.stdout = "700.0\n"
+            result.returncode = 0
+            return result
+        elif "ffmpeg" in cmd:
+            # Create an empty temp file (simulate chunk extraction)
+            # The output file is the last argument
+            import pathlib
+            output_file = cmd[-1]
+            pathlib.Path(output_file).touch()
+            result = MagicMock()
+            result.returncode = 0
+            return result
+        return original_run(cmd, **kwargs)
+
+    # Mock WhisperModel
+    mock_model = MagicMock()
+    def mock_transcribe_call(file_path, **kwargs):
+        mock_segments = []
+        for i in range(3):
+            seg = MagicMock()
+            seg.start = i * 1.0
+            seg.end = (i + 1) * 1.0
+            seg.text = f"Segment {i}"
+            seg.words = []
+            mock_segments.append(seg)
+        mock_info = MagicMock()
+        mock_info.language = "en"
+        mock_info.language_probability = 0.99
+        mock_info.duration = 300.0
+        return iter(mock_segments), mock_info
+
+    mock_model.transcribe = mock_transcribe_call
+
+    service = TranscribeService()
+    service._model = mock_model
+    service._current_model_name = "base"
+    service._current_device = "cpu"
+    service._current_compute_type = "int8"
+
+    written_messages = []
+    def mock_write(msg):
+        written_messages.append(msg)
+
+    with patch("subprocess.run", mock_subprocess_run), \
+         patch("voice_to_notes.services.transcribe.write_message", mock_write):
+        result = service.transcribe_chunked("req-1", "/fake/long_audio.wav")
+
+    # Should have segments from multiple chunks
+    assert len(result.segments) > 0
+
+    # Verify timestamp offsets — segments from chunk 1 should start at 0,
+    # segments from chunk 2 should be offset by 300000ms
+    if len(result.segments) > 3:
+        # Chunk 2 segments should have offset timestamps
+        assert result.segments[3].start_ms >= 300000
+
+    assert result.duration_ms == 700000
+    assert result.language == "en"

python/voice_to_notes.spec

@@ -0,0 +1,67 @@
+# -*- mode: python ; coding: utf-8 -*-
+"""PyInstaller spec for the Voice to Notes sidecar binary."""
+
+from PyInstaller.utils.hooks import collect_all
+
+block_cipher = None
+
+# Collect all files for packages that have shared libraries / data files
+# PyInstaller often misses these for ML packages
+ctranslate2_datas, ctranslate2_binaries, ctranslate2_hiddenimports = collect_all("ctranslate2")
+faster_whisper_datas, faster_whisper_binaries, faster_whisper_hiddenimports = collect_all(
+    "faster_whisper"
+)
+pyannote_datas, pyannote_binaries, pyannote_hiddenimports = collect_all("pyannote")
+
+a = Analysis(
+    ["voice_to_notes/main.py"],
+    pathex=[],
+    binaries=ctranslate2_binaries + faster_whisper_binaries + pyannote_binaries,
+    datas=ctranslate2_datas + faster_whisper_datas + pyannote_datas,
+    hiddenimports=[
+        "torch",
+        "torchaudio",
+        "huggingface_hub",
+        "pysubs2",
+        "openai",
+        "anthropic",
+        "litellm",
+    ]
+    + ctranslate2_hiddenimports
+    + faster_whisper_hiddenimports
+    + pyannote_hiddenimports,
+    hookspath=[],
+    hooksconfig={},
+    runtime_hooks=[],
+    excludes=["tkinter", "test", "unittest", "pip", "setuptools"],
+    win_no_prefer_redirects=False,
+    win_private_assemblies=False,
+    cipher=block_cipher,
+    noarchive=False,
+)
+
+pyz = PYZ(a.pure, a.zipped_data, cipher=block_cipher)
+
+exe = EXE(
+    pyz,
+    a.scripts,
+    [],
+    exclude_binaries=True,
+    name="voice-to-notes-sidecar",
+    debug=False,
+    bootloader_ignore_signals=False,
+    strip=False,
+    upx=True,
+    console=True,
+)
+
+coll = COLLECT(
+    exe,
+    a.binaries,
+    a.zipfiles,
+    a.datas,
+    strip=False,
+    upx=True,
+    upx_exclude=[],
+    name="voice-to-notes-sidecar",
+)

python/voice_to_notes/hardware/detect.py

@@ -2,7 +2,10 @@
 
 from __future__ import annotations
 
+import ctypes
 import os
+import platform
+import subprocess
 import sys
 from dataclasses import dataclass
 
@@ -21,6 +24,77 @@ class HardwareInfo:
     recommended_compute_type: str = "int8"
 
 
+def _detect_ram_mb() -> int:
+    """Detect total system RAM in MB (cross-platform).
+
+    Tries platform-specific methods in order:
+    1. Linux: read /proc/meminfo
+    2. macOS: sysctl hw.memsize
+    3. Windows: GlobalMemoryStatusEx via ctypes
+    4. Fallback: os.sysconf (most Unix systems)
+
+    Returns 0 if all methods fail.
+    """
+    # Linux: read /proc/meminfo
+    if sys.platform == "linux":
+        try:
+            with open("/proc/meminfo") as f:
+                for line in f:
+                    if line.startswith("MemTotal:"):
+                        # Value is in kB
+                        return int(line.split()[1]) // 1024
+        except (FileNotFoundError, ValueError, OSError):
+            pass
+
+    # macOS: sysctl hw.memsize (returns bytes)
+    if sys.platform == "darwin":
+        try:
+            result = subprocess.run(
+                ["sysctl", "-n", "hw.memsize"],
+                capture_output=True,
+                text=True,
+                check=True,
+            )
+            return int(result.stdout.strip()) // (1024 * 1024)
+        except (subprocess.SubprocessError, ValueError, OSError):
+            pass
+
+    # Windows: GlobalMemoryStatusEx via ctypes
+    if sys.platform == "win32":
+        try:
+
+            class MEMORYSTATUSEX(ctypes.Structure):
+                _fields_ = [
+                    ("dwLength", ctypes.c_ulong),
+                    ("dwMemoryLoad", ctypes.c_ulong),
+                    ("ullTotalPhys", ctypes.c_ulonglong),
+                    ("ullAvailPhys", ctypes.c_ulonglong),
+                    ("ullTotalPageFile", ctypes.c_ulonglong),
+                    ("ullAvailPageFile", ctypes.c_ulonglong),
+                    ("ullTotalVirtual", ctypes.c_ulonglong),
+                    ("ullAvailVirtual", ctypes.c_ulonglong),
+                    ("ullAvailExtendedVirtual", ctypes.c_ulonglong),
+                ]
+
+            mem_status = MEMORYSTATUSEX()
+            mem_status.dwLength = ctypes.sizeof(MEMORYSTATUSEX)
+            if ctypes.windll.kernel32.GlobalMemoryStatusEx(ctypes.byref(mem_status)):
+                return int(mem_status.ullTotalPhys) // (1024 * 1024)
+        except (AttributeError, OSError):
+            pass
+
+    # Fallback: os.sysconf (works on most Unix systems)
+    try:
+        page_size = os.sysconf("SC_PAGE_SIZE")
+        phys_pages = os.sysconf("SC_PHYS_PAGES")
+        if page_size > 0 and phys_pages > 0:
+            return (page_size * phys_pages) // (1024 * 1024)
+    except (ValueError, OSError, AttributeError):
+        pass
+
+    return 0
+
+
 def detect_hardware() -> HardwareInfo:
     """Detect available hardware and recommend model configuration."""
     info = HardwareInfo()
@@ -28,16 +102,8 @@ def detect_hardware() -> HardwareInfo:
     # CPU info
     info.cpu_cores = os.cpu_count() or 1
 
-    # RAM info
-    try:
-        with open("/proc/meminfo") as f:
-            for line in f:
-                if line.startswith("MemTotal:"):
-                    # Value is in kB
-                    info.ram_mb = int(line.split()[1]) // 1024
-                    break
-    except (FileNotFoundError, ValueError):
-        pass
+    # RAM info (cross-platform)
+    info.ram_mb = _detect_ram_mb()
 
     # CUDA detection
     try:

python/voice_to_notes/ipc/handlers.py

@@ -88,6 +88,79 @@ def make_diarize_handler() -> HandlerFunc:
     return handler
 
 
+def make_diarize_download_handler() -> HandlerFunc:
+    """Create a handler that downloads/validates the diarization model."""
+    import os
+
+    def handler(msg: IPCMessage) -> IPCMessage:
+        payload = msg.payload
+        hf_token = payload.get("hf_token")
+
+        try:
+            import huggingface_hub
+
+            # Disable pyannote telemetry (has a bug in v4.0.4)
+            os.environ.setdefault("PYANNOTE_METRICS_ENABLED", "false")
+            from pyannote.audio import Pipeline
+
+            # Persist token globally so ALL huggingface_hub downloads use auth.
+            # Setting env var alone isn't enough — pyannote's internal sub-downloads
+            # (e.g. PLDA.from_pretrained) don't forward the token= parameter.
+            # login() writes the token to ~/.cache/huggingface/token which
+            # huggingface_hub reads automatically for all downloads.
+            if hf_token:
+                os.environ["HF_TOKEN"] = hf_token
+                huggingface_hub.login(token=hf_token, add_to_git_credential=False)
+
+            # Pre-download sub-models that pyannote loads internally.
+            # This ensures they're cached before Pipeline.from_pretrained
+            # tries to load them (where token forwarding can fail).
+            sub_models = [
+                "pyannote/segmentation-3.0",
+                "pyannote/speaker-diarization-community-1",
+            ]
+            for model_id in sub_models:
+                print(f"[sidecar] Pre-downloading {model_id}...", file=sys.stderr, flush=True)
+                huggingface_hub.snapshot_download(model_id, token=hf_token)
+
+            print("[sidecar] Downloading diarization pipeline...", file=sys.stderr, flush=True)
+            pipeline = Pipeline.from_pretrained(
+                "pyannote/speaker-diarization-3.1",
+                token=hf_token,
+            )
+            print("[sidecar] Diarization model downloaded successfully", file=sys.stderr, flush=True)
+            return IPCMessage(
+                id=msg.id,
+                type="diarize.download.result",
+                payload={"ok": True},
+            )
+        except Exception as e:
+            error_msg = str(e)
+            print(f"[sidecar] Model download error: {error_msg}", file=sys.stderr, flush=True)
+            # Make common errors more user-friendly
+            if "403" in error_msg or "gated" in error_msg.lower():
+                # Try to extract the specific model name from the error
+                import re
+                model_match = re.search(r"pyannote/[\w-]+", error_msg)
+                if model_match:
+                    model_name = model_match.group(0)
+                    error_msg = (
+                        f"Access denied for {model_name}. "
+                        f"Please visit huggingface.co/{model_name} "
+                        f"and accept the license agreement, then try again."
+                    )
+                else:
+                    error_msg = (
+                        "Access denied. Please accept the license agreements for all "
+                        "required pyannote models on HuggingFace."
+                    )
+            elif "401" in error_msg:
+                error_msg = "Invalid token. Please check your HuggingFace token."
+            return error_message(msg.id, "download_error", error_msg)
+
+    return handler
+
+
 def make_pipeline_handler() -> HandlerFunc:
     """Create a full pipeline handler (transcribe + diarize + merge)."""
     from voice_to_notes.services.pipeline import PipelineService, pipeline_result_to_payload
@@ -107,6 +180,7 @@ def make_pipeline_handler() -> HandlerFunc:
             min_speakers=payload.get("min_speakers"),
             max_speakers=payload.get("max_speakers"),
             skip_diarization=payload.get("skip_diarization", False),
+            hf_token=payload.get("hf_token"),
         )
         return IPCMessage(
             id=msg.id,
@@ -186,10 +260,12 @@ def make_ai_chat_handler() -> HandlerFunc:
                     model=config.get("model", "claude-sonnet-4-6"),
                 ))
             elif provider_name == "litellm":
-                from voice_to_notes.providers.litellm_provider import LiteLLMProvider
+                from voice_to_notes.providers.litellm_provider import OpenAICompatibleProvider
 
-                service.register_provider("litellm", LiteLLMProvider(
+                service.register_provider("litellm", OpenAICompatibleProvider(
                     model=config.get("model", "gpt-4o-mini"),
+                    api_key=config.get("api_key"),
+                    api_base=config.get("api_base"),
                 ))
             return IPCMessage(
                 id=msg.id,

python/voice_to_notes/ipc/messages.py

@@ -34,6 +34,14 @@ def progress_message(request_id: str, percent: int, stage: str, message: str) ->
     )
 
 
+def partial_segment_message(request_id: str, segment_data: dict) -> IPCMessage:
+    return IPCMessage(id=request_id, type="pipeline.segment", payload=segment_data)
+
+
+def speaker_update_message(request_id: str, updates: list[dict]) -> IPCMessage:
+    return IPCMessage(id=request_id, type="pipeline.speaker_update", payload={"updates": updates})
+
+
 def error_message(request_id: str, code: str, message: str) -> IPCMessage:
     return IPCMessage(
         id=request_id,

python/voice_to_notes/ipc/protocol.py

@@ -1,13 +1,53 @@
-"""JSON-line protocol reader/writer over stdin/stdout."""
+"""JSON-line protocol reader/writer over stdin/stdout.
+
+IMPORTANT: stdout is reserved exclusively for IPC messages.
+At init time we save the real stdout, then redirect sys.stdout → stderr
+so that any rogue print() calls from libraries don't corrupt the IPC stream.
+"""
 
 from __future__ import annotations
 
+import io
 import json
+import os
 import sys
 from typing import Any
 
 from voice_to_notes.ipc.messages import IPCMessage
 
+# Save the real stdout fd for IPC before any library can pollute it.
+# Then redirect sys.stdout to stderr so library prints go to stderr.
+_ipc_out: io.TextIOWrapper | None = None
+
+
+def init_ipc() -> None:
+    """Capture real stdout for IPC and redirect sys.stdout to stderr.
+
+    Must be called once at sidecar startup, before importing any ML libraries.
+    """
+    global _ipc_out
+    if _ipc_out is not None:
+        return  # already initialised
+
+    # Duplicate the real stdout fd so we keep it even after redirect
+    real_stdout_fd = os.dup(sys.stdout.fileno())
+    _ipc_out = io.TextIOWrapper(
+        io.BufferedWriter(io.FileIO(real_stdout_fd, "w")),
+        encoding="utf-8",
+        line_buffering=True,
+    )
+
+    # Redirect sys.stdout → stderr so print() from libraries goes to stderr
+    sys.stdout = sys.stderr
+
+
+def _get_ipc_out() -> io.TextIOWrapper:
+    """Return the IPC output stream, falling back to sys.__stdout__."""
+    if _ipc_out is not None:
+        return _ipc_out
+    # Fallback if init_ipc() was never called (e.g. in tests)
+    return sys.__stdout__
+
 
 def read_message() -> IPCMessage | None:
     """Read a single JSON-line message from stdin. Returns None on EOF."""
@@ -29,17 +69,19 @@ def read_message() -> IPCMessage | None:
 
 
 def write_message(msg: IPCMessage) -> None:
-    """Write a JSON-line message to stdout."""
+    """Write a JSON-line message to the IPC channel (real stdout)."""
+    out = _get_ipc_out()
     line = json.dumps(msg.to_dict(), separators=(",", ":"))
-    sys.stdout.write(line + "\n")
-    sys.stdout.flush()
+    out.write(line + "\n")
+    out.flush()
 
 
 def write_dict(data: dict[str, Any]) -> None:
-    """Write a raw dict as a JSON-line message to stdout."""
+    """Write a raw dict as a JSON-line message to the IPC channel."""
+    out = _get_ipc_out()
     line = json.dumps(data, separators=(",", ":"))
-    sys.stdout.write(line + "\n")
-    sys.stdout.flush()
+    out.write(line + "\n")
+    out.flush()
 
 
 def _log(message: str) -> None:

python/voice_to_notes/main.py

@@ -5,18 +5,25 @@ from __future__ import annotations
 import signal
 import sys
 
-from voice_to_notes.ipc.handlers import (
+# CRITICAL: Capture real stdout for IPC *before* importing any ML libraries
+# that might print to stdout and corrupt the JSON-line protocol.
+from voice_to_notes.ipc.protocol import init_ipc
+
+init_ipc()
+
+from voice_to_notes.ipc.handlers import (  # noqa: E402
     HandlerRegistry,
     hardware_detect_handler,
     make_ai_chat_handler,
+    make_diarize_download_handler,
     make_diarize_handler,
     make_export_handler,
     make_pipeline_handler,
     make_transcribe_handler,
     ping_handler,
 )
-from voice_to_notes.ipc.messages import ready_message
-from voice_to_notes.ipc.protocol import read_message, write_message
+from voice_to_notes.ipc.messages import ready_message  # noqa: E402
+from voice_to_notes.ipc.protocol import read_message, write_message  # noqa: E402
 
 
 def create_registry() -> HandlerRegistry:
@@ -26,6 +33,7 @@ def create_registry() -> HandlerRegistry:
     registry.register("transcribe.start", make_transcribe_handler())
     registry.register("hardware.detect", hardware_detect_handler)
     registry.register("diarize.start", make_diarize_handler())
+    registry.register("diarize.download", make_diarize_download_handler())
     registry.register("pipeline.start", make_pipeline_handler())
     registry.register("export.start", make_export_handler())
     registry.register("ai.chat", make_ai_chat_handler())

python/voice_to_notes/providers/litellm_provider.py

@@ -1,4 +1,4 @@
-"""LiteLLM provider — multi-provider gateway."""
+"""OpenAI-compatible provider — works with any OpenAI-compatible API endpoint."""
 
 from __future__ import annotations
 
@@ -7,36 +7,44 @@ from typing import Any
 from voice_to_notes.providers.base import AIProvider
 
 
-class LiteLLMProvider(AIProvider):
-    """Routes through LiteLLM for access to 100+ LLM providers."""
+class OpenAICompatibleProvider(AIProvider):
+    """Connects to any OpenAI-compatible API (LiteLLM proxy, Ollama, vLLM, etc.)."""
 
-    def __init__(self, model: str = "gpt-4o-mini", **kwargs: Any) -> None:
+    def __init__(
+        self,
+        api_key: str | None = None,
+        api_base: str | None = None,
+        model: str = "gpt-4o-mini",
+        **kwargs: Any,
+    ) -> None:
+        self._api_key = api_key or "sk-no-key"
+        self._api_base = api_base
         self._model = model
         self._extra_kwargs = kwargs
 
     def chat(self, messages: list[dict[str, str]], **kwargs: Any) -> str:
-        try:
-            import litellm
-        except ImportError:
-            raise RuntimeError("litellm package is required. Install with: pip install litellm")
+        from openai import OpenAI
 
-        merged_kwargs = {**self._extra_kwargs, **kwargs}
-        response = litellm.completion(
-            model=merged_kwargs.get("model", self._model),
+        client_kwargs: dict[str, Any] = {"api_key": self._api_key}
+        if self._api_base:
+            client_kwargs["base_url"] = self._api_base
+
+        client = OpenAI(**client_kwargs)
+        response = client.chat.completions.create(
+            model=kwargs.get("model", self._model),
             messages=messages,
-            temperature=merged_kwargs.get("temperature", 0.7),
-            max_tokens=merged_kwargs.get("max_tokens", 2048),
+            temperature=kwargs.get("temperature", 0.7),
+            max_tokens=kwargs.get("max_tokens", 2048),
         )
         return response.choices[0].message.content or ""
 
     def is_available(self) -> bool:
         try:
-            import litellm  # noqa: F401
-
-            return True
+            import openai  # noqa: F401
+            return bool(self._api_key and self._api_base)
         except ImportError:
             return False
 
     @property
     def name(self) -> str:
-        return "LiteLLM"
+        return "OpenAI Compatible"

python/voice_to_notes/services/ai_provider.py

@@ -92,7 +92,7 @@ class AIProviderService:
 def create_default_service() -> AIProviderService:
     """Create an AIProviderService with all supported providers registered."""
     from voice_to_notes.providers.anthropic_provider import AnthropicProvider
-    from voice_to_notes.providers.litellm_provider import LiteLLMProvider
+    from voice_to_notes.providers.litellm_provider import OpenAICompatibleProvider
     from voice_to_notes.providers.local_provider import LocalProvider
     from voice_to_notes.providers.openai_provider import OpenAIProvider
 
@@ -100,5 +100,5 @@ def create_default_service() -> AIProviderService:
     service.register_provider("local", LocalProvider())
     service.register_provider("openai", OpenAIProvider())
     service.register_provider("anthropic", AnthropicProvider())
-    service.register_provider("litellm", LiteLLMProvider())
+    service.register_provider("litellm", OpenAICompatibleProvider())
     return service

python/voice_to_notes/services/diarize.py

@@ -2,15 +2,69 @@
 
 from __future__ import annotations
 
+import os
+import subprocess
 import sys
+import tempfile
+import threading
 import time
 from dataclasses import dataclass, field
+from pathlib import Path
 from typing import Any
 
+# Disable pyannote telemetry — it has a bug in v4.0.4 where
+# np.isfinite(None) crashes when max_speakers is not set.
+os.environ.setdefault("PYANNOTE_METRICS_ENABLED", "false")
+
+from voice_to_notes.utils.ffmpeg import get_ffmpeg_path
 from voice_to_notes.ipc.messages import progress_message
 from voice_to_notes.ipc.protocol import write_message
 
 
+def _ensure_wav(file_path: str) -> tuple[str, str | None]:
+    """Convert audio to 16kHz mono WAV if needed.
+
+    pyannote.audio v4.0.4 has a bug where its AudioDecoder returns
+    duration=None for some formats (FLAC, etc.), causing crashes.
+    Converting to WAV ensures the duration header is always present.
+
+    Returns:
+        (path_to_use, temp_path_or_None)
+        If conversion was needed, temp_path is the WAV file to clean up.
+    """
+    ext = Path(file_path).suffix.lower()
+    if ext == ".wav":
+        return file_path, None
+
+    tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
+    tmp.close()
+    try:
+        subprocess.run(
+            [
+                get_ffmpeg_path(), "-y", "-i", file_path,
+                "-ar", "16000", "-ac", "1", "-c:a", "pcm_s16le",
+                tmp.name,
+            ],
+            check=True,
+            capture_output=True,
+        )
+        print(
+            f"[sidecar] Converted {ext} to WAV for diarization",
+            file=sys.stderr,
+            flush=True,
+        )
+        return tmp.name, tmp.name
+    except (subprocess.CalledProcessError, FileNotFoundError) as e:
+        # ffmpeg not available or failed — try original file and hope for the best
+        print(
+            f"[sidecar] WAV conversion failed ({e}), using original file",
+            file=sys.stderr,
+            flush=True,
+        )
+        os.unlink(tmp.name)
+        return file_path, None
+
+
 @dataclass
 class SpeakerSegment:
     """A time span assigned to a speaker."""
@@ -35,45 +89,59 @@ class DiarizeService:
     def __init__(self) -> None:
         self._pipeline: Any = None
 
-    def _ensure_pipeline(self) -> Any:
+    def _ensure_pipeline(self, hf_token: str | None = None) -> Any:
         """Load the pyannote diarization pipeline (lazy)."""
         if self._pipeline is not None:
             return self._pipeline
 
         print("[sidecar] Loading pyannote diarization pipeline...", file=sys.stderr, flush=True)
 
-        try:
-            from pyannote.audio import Pipeline
+        # Use token from argument, fall back to environment variable
+        if not hf_token:
+            hf_token = os.environ.get("HF_TOKEN") or os.environ.get("HUGGING_FACE_HUB_TOKEN") or None
 
-            self._pipeline = Pipeline.from_pretrained(
-                "pyannote/speaker-diarization-3.1",
-                use_auth_token=False,
-            )
-        except Exception:
-            # Fall back to a simpler approach if the model isn't available
-            # pyannote requires HuggingFace token for some models
-            # Try the community model first
+        # Persist token globally so ALL huggingface_hub sub-downloads use auth.
+        # Pyannote has internal dependencies that don't forward the token= param.
+        if hf_token:
+            os.environ["HF_TOKEN"] = hf_token
+            import huggingface_hub
+            huggingface_hub.login(token=hf_token, add_to_git_credential=False)
+
+        models = [
+            "pyannote/speaker-diarization-3.1",
+            "pyannote/speaker-diarization",
+        ]
+
+        last_error: Exception | None = None
+        for model_name in models:
             try:
                 from pyannote.audio import Pipeline
 
-                self._pipeline = Pipeline.from_pretrained(
-                    "pyannote/speaker-diarization",
-                    use_auth_token=False,
-                )
+                self._pipeline = Pipeline.from_pretrained(model_name, token=hf_token)
+                print(f"[sidecar] Loaded diarization model: {model_name}", file=sys.stderr, flush=True)
+                # Move pipeline to GPU if available
+                try:
+                    import torch
+                    if torch.cuda.is_available():
+                        self._pipeline = self._pipeline.to(torch.device("cuda"))
+                        print(f"[sidecar] Diarization pipeline moved to GPU", file=sys.stderr, flush=True)
+                except Exception as e:
+                    print(f"[sidecar] GPU not available for diarization: {e}", file=sys.stderr, flush=True)
+                return self._pipeline
             except Exception as e:
+                last_error = e
                 print(
-                    f"[sidecar] Warning: Could not load pyannote pipeline: {e}",
+                    f"[sidecar] Warning: Could not load {model_name}: {e}",
                     file=sys.stderr,
                     flush=True,
                 )
-                raise RuntimeError(
-                    "pyannote.audio pipeline not available. "
-                    "You may need to accept the model license at "
-                    "https://huggingface.co/pyannote/speaker-diarization-3.1 "
-                    "and set a HF_TOKEN environment variable."
-                ) from e
 
-        return self._pipeline
+        raise RuntimeError(
+            "pyannote.audio pipeline not available. "
+            "You may need to accept the model license at "
+            "https://huggingface.co/pyannote/speaker-diarization-3.1 "
+            "and set a HF_TOKEN environment variable."
+        ) from last_error
 
     def diarize(
         self,
@@ -82,6 +150,8 @@ class DiarizeService:
         num_speakers: int | None = None,
         min_speakers: int | None = None,
         max_speakers: int | None = None,
+        hf_token: str | None = None,
+        audio_duration_sec: float | None = None,
     ) -> DiarizationResult:
         """Run speaker diarization on an audio file.
 
@@ -99,7 +169,7 @@ class DiarizeService:
             progress_message(request_id, 0, "loading_diarization", "Loading diarization model...")
         )
 
-        pipeline = self._ensure_pipeline()
+        pipeline = self._ensure_pipeline(hf_token=hf_token)
 
         write_message(
             progress_message(request_id, 20, "diarizing", "Running speaker diarization...")
@@ -116,8 +186,55 @@ class DiarizeService:
         if max_speakers is not None:
             kwargs["max_speakers"] = max_speakers
 
-        # Run diarization
-        diarization = pipeline(file_path, **kwargs)
+        # Convert to WAV to work around pyannote v4.0.4 duration bug
+        audio_path, temp_wav = _ensure_wav(file_path)
+
+        print(
+            f"[sidecar] Running diarization on {audio_path} with kwargs: {kwargs}",
+            file=sys.stderr,
+            flush=True,
+        )
+
+        # Run diarization in background thread for progress reporting
+        result_holder: list = [None]
+        error_holder: list[Exception | None] = [None]
+        done_event = threading.Event()
+
+        def _run():
+            try:
+                result_holder[0] = pipeline(audio_path, **kwargs)
+            except Exception as e:
+                error_holder[0] = e
+            finally:
+                done_event.set()
+
+        thread = threading.Thread(target=_run, daemon=True)
+        thread.start()
+
+        elapsed = 0.0
+        estimated_total = max(audio_duration_sec * 0.5, 30.0) if audio_duration_sec else 120.0
+        while not done_event.wait(timeout=2.0):
+            elapsed += 2.0
+            pct = min(20 + int((elapsed / estimated_total) * 65), 85)
+            write_message(progress_message(
+                request_id, pct, "diarizing",
+                f"Analyzing speakers ({int(elapsed)}s elapsed)..."))
+
+        thread.join()
+
+        # Clean up temp file
+        if temp_wav:
+            os.unlink(temp_wav)
+
+        if error_holder[0] is not None:
+            raise error_holder[0]
+        raw_result = result_holder[0]
+
+        # pyannote 4.0+ returns DiarizeOutput; older versions return Annotation directly
+        if hasattr(raw_result, "speaker_diarization"):
+            diarization = raw_result.speaker_diarization
+        else:
+            diarization = raw_result
 
         # Convert pyannote output to our format
         result = DiarizationResult()

python/voice_to_notes/services/pipeline.py

@@ -2,13 +2,19 @@
 
 from __future__ import annotations
 
+import concurrent.futures
 import sys
 import time
 from dataclasses import dataclass, field
 from typing import Any
 
-from voice_to_notes.ipc.messages import progress_message
+from voice_to_notes.ipc.messages import (
+    partial_segment_message,
+    progress_message,
+    speaker_update_message,
+)
 from voice_to_notes.ipc.protocol import write_message
+from voice_to_notes.utils.ffmpeg import get_ffprobe_path
 from voice_to_notes.services.diarize import DiarizeService, SpeakerSegment
 from voice_to_notes.services.transcribe import (
     SegmentResult,
@@ -60,6 +66,7 @@ class PipelineService:
         min_speakers: int | None = None,
         max_speakers: int | None = None,
         skip_diarization: bool = False,
+        hf_token: str | None = None,
     ) -> PipelineResult:
         """Run the full transcription + diarization pipeline.
 
@@ -77,22 +84,59 @@ class PipelineService:
         """
         start_time = time.time()
 
-        # Step 1: Transcribe
+        # Step 0: Probe audio duration for conditional chunked transcription
         write_message(
             progress_message(request_id, 0, "pipeline", "Starting transcription pipeline...")
         )
 
-        transcription = self._transcribe_service.transcribe(
-            request_id=request_id,
-            file_path=file_path,
-            model_name=model_name,
-            device=device,
-            compute_type=compute_type,
-            language=language,
-        )
+        def _emit_segment(seg: SegmentResult, index: int) -> None:
+            write_message(partial_segment_message(request_id, {
+                "index": index,
+                "text": seg.text,
+                "start_ms": seg.start_ms,
+                "end_ms": seg.end_ms,
+                "words": [{"word": w.word, "start_ms": w.start_ms, "end_ms": w.end_ms, "confidence": w.confidence} for w in seg.words],
+            }))
+
+        audio_duration_sec = None
+        try:
+            import subprocess
+            probe_result = subprocess.run(
+                [get_ffprobe_path(), "-v", "quiet", "-show_entries", "format=duration",
+                 "-of", "default=noprint_wrappers=1:nokey=1", file_path],
+                capture_output=True, text=True, check=True,
+            )
+            audio_duration_sec = float(probe_result.stdout.strip())
+        except (subprocess.CalledProcessError, FileNotFoundError, ValueError):
+            pass
+
+        def _run_transcription() -> TranscriptionResult:
+            """Run transcription (chunked or standard based on duration)."""
+            from voice_to_notes.services.transcribe import LARGE_FILE_THRESHOLD_SEC
+            if audio_duration_sec and audio_duration_sec > LARGE_FILE_THRESHOLD_SEC:
+                return self._transcribe_service.transcribe_chunked(
+                    request_id=request_id,
+                    file_path=file_path,
+                    model_name=model_name,
+                    device=device,
+                    compute_type=compute_type,
+                    language=language,
+                    on_segment=_emit_segment,
+                )
+            else:
+                return self._transcribe_service.transcribe(
+                    request_id=request_id,
+                    file_path=file_path,
+                    model_name=model_name,
+                    device=device,
+                    compute_type=compute_type,
+                    language=language,
+                    on_segment=_emit_segment,
+                )
 
         if skip_diarization:
-            # Convert transcription directly without speaker labels
+            # Sequential: transcribe only, no diarization needed
+            transcription = _run_transcription()
             result = PipelineResult(
                 language=transcription.language,
                 language_probability=transcription.language_probability,
@@ -110,27 +154,83 @@ class PipelineService:
                 )
             return result
 
-        # Step 2: Diarize
+        # Parallel execution: run transcription (0-45%) and diarization (45-90%)
+        # concurrently, then merge (90-100%).
         write_message(
-            progress_message(request_id, 50, "pipeline", "Starting speaker diarization...")
+            progress_message(
+                request_id, 0, "pipeline",
+                "Starting transcription and diarization in parallel..."
+            )
         )
 
-        diarization = self._diarize_service.diarize(
-            request_id=request_id,
-            file_path=file_path,
-            num_speakers=num_speakers,
-            min_speakers=min_speakers,
-            max_speakers=max_speakers,
-        )
+        diarization = None
+        diarization_error = None
 
-        # Step 3: Merge
-        write_message(
-            progress_message(request_id, 90, "pipeline", "Merging transcript with speakers...")
-        )
+        with concurrent.futures.ThreadPoolExecutor(max_workers=2) as executor:
+            transcription_future = executor.submit(_run_transcription)
 
-        result = self._merge_results(transcription, diarization.speaker_segments)
-        result.speakers = diarization.speakers
-        result.num_speakers = diarization.num_speakers
+            # Use probed audio_duration_sec for diarization progress estimation
+            # (transcription hasn't finished yet, so we can't use transcription.duration_ms)
+            diarization_future = executor.submit(
+                self._diarize_service.diarize,
+                request_id=request_id,
+                file_path=file_path,
+                num_speakers=num_speakers,
+                min_speakers=min_speakers,
+                max_speakers=max_speakers,
+                hf_token=hf_token,
+                audio_duration_sec=audio_duration_sec,
+            )
+
+            # Wait for both futures. We need the transcription result regardless,
+            # but diarization may fail gracefully.
+            transcription = transcription_future.result()
+            write_message(
+                progress_message(request_id, 45, "pipeline", "Transcription complete")
+            )
+
+            try:
+                diarization = diarization_future.result()
+            except Exception as e:
+                import traceback
+                diarization_error = e
+                print(
+                    f"[sidecar] Diarization failed, falling back to transcription-only: {e}",
+                    file=sys.stderr,
+                    flush=True,
+                )
+                traceback.print_exc(file=sys.stderr)
+                write_message(
+                    progress_message(
+                        request_id, 80, "pipeline",
+                        f"Diarization failed ({e}), using transcription only..."
+                    )
+                )
+
+        # Step 3: Merge (or skip if diarization failed)
+        if diarization is not None:
+            write_message(
+                progress_message(request_id, 90, "merging", "Merging transcript with speakers...")
+            )
+            result = self._merge_results(transcription, diarization.speaker_segments)
+            result.speakers = diarization.speakers
+            result.num_speakers = diarization.num_speakers
+        else:
+            result = PipelineResult(
+                language=transcription.language,
+                language_probability=transcription.language_probability,
+                duration_ms=transcription.duration_ms,
+            )
+            for seg in transcription.segments:
+                result.segments.append(
+                    PipelineSegment(
+                        text=seg.text,
+                        start_ms=seg.start_ms,
+                        end_ms=seg.end_ms,
+                        speaker=None,
+                        words=seg.words,
+                    )
+                )
 
         elapsed = time.time() - start_time
         print(
@@ -140,6 +240,10 @@ class PipelineService:
             flush=True,
         )
 
+        updates = [{"index": i, "speaker": seg.speaker} for i, seg in enumerate(result.segments) if seg.speaker]
+        if updates:
+            write_message(speaker_update_message(request_id, updates))
+
         write_message(
             progress_message(request_id, 100, "done", "Pipeline complete")
         )

python/voice_to_notes/services/transcribe.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 import sys
 import time
+from collections.abc import Callable
 from dataclasses import dataclass, field
 from typing import Any
 
@@ -11,6 +12,10 @@ from faster_whisper import WhisperModel
 
 from voice_to_notes.ipc.messages import progress_message
 from voice_to_notes.ipc.protocol import write_message
+from voice_to_notes.utils.ffmpeg import get_ffmpeg_path, get_ffprobe_path
+
+CHUNK_REPORT_SIZE = 10
+LARGE_FILE_THRESHOLD_SEC = 3600  # 1 hour
 
 
 @dataclass
@@ -90,6 +95,7 @@ class TranscribeService:
         device: str = "cpu",
         compute_type: str = "int8",
         language: str | None = None,
+        on_segment: Callable[[SegmentResult, int], None] | None = None,
     ) -> TranscriptionResult:
         """Transcribe an audio file with word-level timestamps.
 
@@ -145,16 +151,23 @@ class TranscribeService:
                 )
             )
 
-            # Send progress every few segments
-            if segment_count % 5 == 0:
-                write_message(
-                    progress_message(
-                        request_id,
-                        progress_pct,
-                        "transcribing",
-                        f"Processed {segment_count} segments...",
-                    )
+            if on_segment:
+                on_segment(result.segments[-1], segment_count - 1)
+
+            write_message(
+                progress_message(
+                    request_id,
+                    progress_pct,
+                    "transcribing",
+                    f"Transcribing segment {segment_count} ({progress_pct}% of audio)...",
                 )
+            )
+
+            if segment_count % CHUNK_REPORT_SIZE == 0:
+                write_message(progress_message(
+                    request_id, progress_pct, "transcribing",
+                    f"Completed chunk of {CHUNK_REPORT_SIZE} segments "
+                    f"({segment_count} total, {progress_pct}% of audio)..."))
 
         elapsed = time.time() - start_time
         print(
@@ -166,6 +179,113 @@ class TranscribeService:
         write_message(progress_message(request_id, 100, "done", "Transcription complete"))
         return result
 
+    def transcribe_chunked(
+        self,
+        request_id: str,
+        file_path: str,
+        model_name: str = "base",
+        device: str = "cpu",
+        compute_type: str = "int8",
+        language: str | None = None,
+        on_segment: Callable[[SegmentResult, int], None] | None = None,
+        chunk_duration_sec: int = 300,
+    ) -> TranscriptionResult:
+        """Transcribe a large audio file by splitting into chunks.
+
+        Uses ffmpeg to split the file into chunks, transcribes each chunk,
+        then merges the results with corrected timestamps.
+
+        Falls back to standard transcribe() if ffmpeg is not available.
+        """
+        import subprocess
+        import tempfile
+
+        # Get total duration via ffprobe
+        try:
+            probe_result = subprocess.run(
+                [get_ffprobe_path(), "-v", "quiet", "-show_entries", "format=duration",
+                 "-of", "default=noprint_wrappers=1:nokey=1", file_path],
+                capture_output=True, text=True, check=True,
+            )
+            total_duration = float(probe_result.stdout.strip())
+        except (subprocess.CalledProcessError, FileNotFoundError, ValueError):
+            # ffprobe not available or failed — fall back to standard transcription
+            write_message(progress_message(
+                request_id, 5, "transcribing",
+                "ffmpeg not available, using standard transcription..."))
+            return self.transcribe(request_id, file_path, model_name, device,
+                                   compute_type, language, on_segment=on_segment)
+
+        num_chunks = max(1, int(total_duration / chunk_duration_sec) + 1)
+        write_message(progress_message(
+            request_id, 5, "transcribing",
+            f"Splitting {total_duration:.0f}s file into {num_chunks} chunks..."))
+
+        merged_result = TranscriptionResult()
+        global_segment_index = 0
+
+        for chunk_idx in range(num_chunks):
+            chunk_start = chunk_idx * chunk_duration_sec
+            if chunk_start >= total_duration:
+                break
+
+            chunk_start_ms = int(chunk_start * 1000)
+
+            # Extract chunk to temp file
+            tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
+            tmp.close()
+            try:
+                subprocess.run(
+                    [get_ffmpeg_path(), "-y", "-ss", str(chunk_start),
+                     "-t", str(chunk_duration_sec),
+                     "-i", file_path,
+                     "-ar", "16000", "-ac", "1", "-c:a", "pcm_s16le",
+                     tmp.name],
+                    capture_output=True, check=True,
+                )
+
+                # Wrap on_segment to offset the index
+                chunk_on_segment = None
+                if on_segment:
+                    base_index = global_segment_index
+                    def chunk_on_segment(seg: SegmentResult, idx: int, _base=base_index) -> None:
+                        on_segment(seg, _base + idx)
+
+                chunk_result = self.transcribe(
+                    request_id, tmp.name, model_name, device,
+                    compute_type, language, on_segment=chunk_on_segment,
+                )
+
+                # Offset timestamps and merge
+                for seg in chunk_result.segments:
+                    seg.start_ms += chunk_start_ms
+                    seg.end_ms += chunk_start_ms
+                    for word in seg.words:
+                        word.start_ms += chunk_start_ms
+                        word.end_ms += chunk_start_ms
+                    merged_result.segments.append(seg)
+
+                global_segment_index += len(chunk_result.segments)
+
+                # Take language from first chunk
+                if chunk_idx == 0:
+                    merged_result.language = chunk_result.language
+                    merged_result.language_probability = chunk_result.language_probability
+
+            finally:
+                import os
+                os.unlink(tmp.name)
+
+            # Chunk progress
+            chunk_pct = min(10 + int(((chunk_idx + 1) / num_chunks) * 80), 90)
+            write_message(progress_message(
+                request_id, chunk_pct, "transcribing",
+                f"Completed chunk {chunk_idx + 1}/{num_chunks}..."))
+
+        merged_result.duration_ms = int(total_duration * 1000)
+        write_message(progress_message(request_id, 100, "done", "Transcription complete"))
+        return merged_result
+
 
 def result_to_payload(result: TranscriptionResult) -> dict[str, Any]:
     """Convert TranscriptionResult to IPC payload dict."""

python/voice_to_notes/utils/ffmpeg.py

@@ -0,0 +1,43 @@
+"""Resolve ffmpeg/ffprobe paths for both frozen and development builds."""
+
+from __future__ import annotations
+
+import os
+import sys
+
+
+def get_ffmpeg_path() -> str:
+    """Return the path to the ffmpeg binary.
+
+    When running as a frozen PyInstaller bundle, looks next to sys.executable.
+    Otherwise falls back to the system PATH.
+    """
+    if getattr(sys, "frozen", False):
+        # Frozen PyInstaller bundle — ffmpeg is next to the sidecar binary
+        bundle_dir = os.path.dirname(sys.executable)
+        candidates = [
+            os.path.join(bundle_dir, "ffmpeg.exe" if sys.platform == "win32" else "ffmpeg"),
+            os.path.join(bundle_dir, "ffmpeg"),
+        ]
+        for path in candidates:
+            if os.path.isfile(path):
+                return path
+    return "ffmpeg"
+
+
+def get_ffprobe_path() -> str:
+    """Return the path to the ffprobe binary.
+
+    When running as a frozen PyInstaller bundle, looks next to sys.executable.
+    Otherwise falls back to the system PATH.
+    """
+    if getattr(sys, "frozen", False):
+        bundle_dir = os.path.dirname(sys.executable)
+        candidates = [
+            os.path.join(bundle_dir, "ffprobe.exe" if sys.platform == "win32" else "ffprobe"),
+            os.path.join(bundle_dir, "ffprobe"),
+        ]
+        for path in candidates:
+            if os.path.isfile(path):
+                return path
+    return "ffprobe"

src-tauri/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "voice-to-notes"
-version = "0.1.0"
+version = "0.2.0"
 description = "Voice to Notes — desktop transcription with speaker identification"
 authors = ["Voice to Notes Contributors"]
 license = "MIT"
@@ -20,6 +20,7 @@ serde = { version = "1", features = ["derive"] }
 serde_json = "1"
 rusqlite = { version = "0.31", features = ["bundled"] }
 uuid = { version = "1", features = ["v4", "serde"] }
+zip = { version = "2", default-features = false, features = ["deflate"] }
 thiserror = "1"
 chrono = { version = "0.4", features = ["serde"] }
 tauri-plugin-dialog = "2.6.0"

src-tauri/build.rs

@@ -1,3 +1,21 @@
 fn main() {
+    // Ensure sidecar.zip exists so tauri-build doesn't fail.
+    // CI replaces this placeholder with the real PyInstaller sidecar archive.
+    let zip_path = std::path::Path::new("sidecar.zip");
+    if !zip_path.exists() {
+        // Minimal valid zip (empty archive): end-of-central-directory record
+        let empty_zip: [u8; 22] = [
+            0x50, 0x4b, 0x05, 0x06, // EOCD signature
+            0x00, 0x00, // disk number
+            0x00, 0x00, // disk with central dir
+            0x00, 0x00, // entries on this disk
+            0x00, 0x00, // total entries
+            0x00, 0x00, 0x00, 0x00, // central dir size
+            0x00, 0x00, 0x00, 0x00, // central dir offset
+            0x00, 0x00, // comment length
+        ];
+        std::fs::write(zip_path, empty_zip).expect("Failed to create placeholder sidecar.zip");
+    }
+
     tauri_build::build()
 }

src-tauri/src/commands/ai.rs

@@ -1,19 +1,7 @@
 use serde_json::{json, Value};
 
 use crate::sidecar::messages::IPCMessage;
-use crate::sidecar::SidecarManager;
-
-fn get_sidecar() -> Result<SidecarManager, String> {
-    let python_path = std::env::current_dir()
-        .map_err(|e| e.to_string())?
-        .join("../python")
-        .canonicalize()
-        .map_err(|e| format!("Cannot find python directory: {e}"))?;
-
-    let manager = SidecarManager::new();
-    manager.start(&python_path.to_string_lossy())?;
-    Ok(manager)
-}
+use crate::sidecar::sidecar;
 
 /// Send a chat message to the AI provider via the Python sidecar.
 #[tauri::command]
@@ -22,14 +10,8 @@ pub fn ai_chat(
     transcript_context: Option<String>,
     provider: Option<String>,
 ) -> Result<Value, String> {
-    let manager = get_sidecar()?;
-
-    let request_id = uuid::Uuid::new_v4().to_string();
-    let payload = json!({
-        "action": "chat",
-        "messages": messages,
-        "transcript_context": transcript_context.unwrap_or_default(),
-    });
+    let manager = sidecar();
+    manager.ensure_running()?;
 
     // If a specific provider is requested, set it first
     if let Some(p) = provider {
@@ -41,13 +23,27 @@ pub fn ai_chat(
         let _ = manager.send_and_receive(&set_msg)?;
     }
 
-    let msg = IPCMessage::new(&request_id, "ai.chat", payload);
+    let request_id = uuid::Uuid::new_v4().to_string();
+    let msg = IPCMessage::new(
+        &request_id,
+        "ai.chat",
+        json!({
+            "action": "chat",
+            "messages": messages,
+            "transcript_context": transcript_context.unwrap_or_default(),
+        }),
+    );
+
     let response = manager.send_and_receive(&msg)?;
 
     if response.msg_type == "error" {
         return Err(format!(
             "AI error: {}",
-            response.payload.get("message").and_then(|v| v.as_str()).unwrap_or("unknown")
+            response
+                .payload
+                .get("message")
+                .and_then(|v| v.as_str())
+                .unwrap_or("unknown")
         ));
     }
 
@@ -57,7 +53,8 @@ pub fn ai_chat(
 /// List available AI providers.
 #[tauri::command]
 pub fn ai_list_providers() -> Result<Value, String> {
-    let manager = get_sidecar()?;
+    let manager = sidecar();
+    manager.ensure_running()?;
 
     let request_id = uuid::Uuid::new_v4().to_string();
     let msg = IPCMessage::new(
@@ -73,7 +70,8 @@ pub fn ai_list_providers() -> Result<Value, String> {
 /// Configure an AI provider with API key/settings.
 #[tauri::command]
 pub fn ai_configure(provider: String, config: Value) -> Result<Value, String> {
-    let manager = get_sidecar()?;
+    let manager = sidecar();
+    manager.ensure_running()?;
 
     let request_id = uuid::Uuid::new_v4().to_string();
     let msg = IPCMessage::new(

src-tauri/src/commands/export.rs

@@ -1,7 +1,7 @@
 use serde_json::{json, Value};
 
 use crate::sidecar::messages::IPCMessage;
-use crate::sidecar::SidecarManager;
+use crate::sidecar::sidecar;
 
 /// Export transcript to caption/text format via the Python sidecar.
 #[tauri::command]
@@ -12,16 +12,8 @@ pub fn export_transcript(
     output_path: String,
     title: Option<String>,
 ) -> Result<Value, String> {
-    let python_path = std::env::current_dir()
-        .map_err(|e| e.to_string())?
-        .join("../python")
-        .canonicalize()
-        .map_err(|e| format!("Cannot find python directory: {e}"))?;
-
-    let python_path_str = python_path.to_string_lossy().to_string();
-
-    let manager = SidecarManager::new();
-    manager.start(&python_path_str)?;
+    let manager = sidecar();
+    manager.ensure_running()?;
 
     let request_id = uuid::Uuid::new_v4().to_string();
     let msg = IPCMessage::new(
@@ -41,7 +33,11 @@ pub fn export_transcript(
     if response.msg_type == "error" {
         return Err(format!(
             "Export error: {}",
-            response.payload.get("message").and_then(|v| v.as_str()).unwrap_or("unknown")
+            response
+                .payload
+                .get("message")
+                .and_then(|v| v.as_str())
+                .unwrap_or("unknown")
         ));
     }
 

src-tauri/src/commands/project.rs

@@ -1,27 +1,297 @@
+use serde::{Deserialize, Serialize};
+use std::fs;
+use tauri::State;
+
 use crate::db::models::Project;
+use crate::db::queries;
+use crate::state::AppState;
+
+// ── File-based project types ────────────────────────────────────
+
+#[derive(Serialize, Deserialize)]
+pub struct ProjectFile {
+    pub version: u32,
+    pub name: String,
+    pub audio_file: String,
+    pub created_at: String,
+    pub segments: Vec<ProjectFileSegment>,
+    pub speakers: Vec<ProjectFileSpeaker>,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct ProjectFileSegment {
+    pub text: String,
+    pub start_ms: i64,
+    pub end_ms: i64,
+    pub speaker: Option<String>,
+    pub is_edited: bool,
+    pub words: Vec<ProjectFileWord>,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct ProjectFileWord {
+    pub word: String,
+    pub start_ms: i64,
+    pub end_ms: i64,
+    pub confidence: f64,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct ProjectFileSpeaker {
+    pub label: String,
+    pub display_name: Option<String>,
+    pub color: String,
+}
+
+// ── Input types for save_project_transcript ──────────────────────
+
+#[derive(Deserialize)]
+pub struct WordInput {
+    pub word: String,
+    pub start_ms: i64,
+    pub end_ms: i64,
+    pub confidence: f64,
+}
+
+#[derive(Deserialize)]
+pub struct SegmentInput {
+    pub text: String,
+    pub start_ms: i64,
+    pub end_ms: i64,
+    pub speaker: Option<String>, // speaker label, not id
+    pub words: Vec<WordInput>,
+}
+
+#[derive(Deserialize)]
+pub struct SpeakerInput {
+    pub label: String,
+    pub color: String,
+}
+
+// ── Output types for load_project_transcript ─────────────────────
+
+#[derive(Serialize)]
+pub struct WordOutput {
+    pub word: String,
+    pub start_ms: i64,
+    pub end_ms: i64,
+    pub confidence: Option<f64>,
+}
+
+#[derive(Serialize)]
+pub struct SegmentOutput {
+    pub id: String,
+    pub text: String,
+    pub start_ms: i64,
+    pub end_ms: i64,
+    pub speaker: Option<String>, // speaker label
+    pub words: Vec<WordOutput>,
+}
+
+#[derive(Serialize)]
+pub struct SpeakerOutput {
+    pub id: String,
+    pub label: String,
+    pub display_name: Option<String>,
+    pub color: Option<String>,
+}
+
+#[derive(Serialize)]
+pub struct ProjectTranscript {
+    pub file_path: String,
+    pub segments: Vec<SegmentOutput>,
+    pub speakers: Vec<SpeakerOutput>,
+}
+
+// ── Commands ─────────────────────────────────────────────────────
 
 #[tauri::command]
-pub fn create_project(name: String) -> Result<Project, String> {
-    // TODO: Use actual database connection from app state
-    Ok(Project {
-        id: uuid::Uuid::new_v4().to_string(),
-        name,
-        created_at: chrono::Utc::now().to_rfc3339(),
-        updated_at: chrono::Utc::now().to_rfc3339(),
-        settings: None,
-        status: "active".to_string(),
-    })
+pub fn create_project(name: String, state: State<AppState>) -> Result<Project, String> {
+    let conn = state.db.lock().map_err(|e| e.to_string())?;
+    queries::create_project(&conn, &name).map_err(|e| e.to_string())
 }
 
 #[tauri::command]
-pub fn get_project(id: String) -> Result<Option<Project>, String> {
-    // TODO: Use actual database connection from app state
-    let _ = id;
-    Ok(None)
+pub fn get_project(id: String, state: State<AppState>) -> Result<Option<Project>, String> {
+    let conn = state.db.lock().map_err(|e| e.to_string())?;
+    queries::get_project(&conn, &id).map_err(|e| e.to_string())
 }
 
 #[tauri::command]
-pub fn list_projects() -> Result<Vec<Project>, String> {
-    // TODO: Use actual database connection from app state
-    Ok(vec![])
+pub fn list_projects(state: State<AppState>) -> Result<Vec<Project>, String> {
+    let conn = state.db.lock().map_err(|e| e.to_string())?;
+    queries::list_projects(&conn).map_err(|e| e.to_string())
+}
+
+#[tauri::command]
+pub fn delete_project(id: String, state: State<AppState>) -> Result<(), String> {
+    let conn = state.db.lock().map_err(|e| e.to_string())?;
+    queries::delete_project(&conn, &id).map_err(|e| e.to_string())
+}
+
+#[tauri::command]
+pub fn update_segment(
+    segment_id: String,
+    new_text: String,
+    state: State<AppState>,
+) -> Result<(), String> {
+    let conn = state.db.lock().map_err(|e| e.to_string())?;
+    queries::update_segment_text(&conn, &segment_id, &new_text).map_err(|e| e.to_string())
+}
+
+#[tauri::command]
+pub fn save_project_transcript(
+    project_id: String,
+    file_path: String,
+    segments: Vec<SegmentInput>,
+    speakers: Vec<SpeakerInput>,
+    state: State<AppState>,
+) -> Result<Project, String> {
+    let conn = state.db.lock().map_err(|e| e.to_string())?;
+
+    // 1. Create media file entry
+    let media_file =
+        queries::create_media_file(&conn, &project_id, &file_path).map_err(|e| e.to_string())?;
+
+    // 2. Create speaker entries and build label -> id map
+    let mut speaker_map = std::collections::HashMap::new();
+    for speaker_input in &speakers {
+        let speaker = queries::create_speaker(
+            &conn,
+            &project_id,
+            &speaker_input.label,
+            Some(&speaker_input.color),
+        )
+        .map_err(|e| e.to_string())?;
+        speaker_map.insert(speaker_input.label.clone(), speaker.id);
+    }
+
+    // 3. Create segments with words
+    for (index, seg_input) in segments.iter().enumerate() {
+        let speaker_id = seg_input
+            .speaker
+            .as_ref()
+            .and_then(|label| speaker_map.get(label));
+
+        let segment = queries::create_segment(
+            &conn,
+            &project_id,
+            &media_file.id,
+            speaker_id.map(|s| s.as_str()),
+            seg_input.start_ms,
+            seg_input.end_ms,
+            &seg_input.text,
+            index as i32,
+        )
+        .map_err(|e| e.to_string())?;
+
+        // Create words for this segment
+        for (word_index, word_input) in seg_input.words.iter().enumerate() {
+            queries::create_word(
+                &conn,
+                &segment.id,
+                &word_input.word,
+                word_input.start_ms,
+                word_input.end_ms,
+                Some(word_input.confidence),
+                word_index as i32,
+            )
+            .map_err(|e| e.to_string())?;
+        }
+    }
+
+    // 4. Return updated project info
+    queries::get_project(&conn, &project_id)
+        .map_err(|e| e.to_string())?
+        .ok_or_else(|| "Project not found".to_string())
+}
+
+#[tauri::command]
+pub fn load_project_transcript(
+    project_id: String,
+    state: State<AppState>,
+) -> Result<Option<ProjectTranscript>, String> {
+    let conn = state.db.lock().map_err(|e| e.to_string())?;
+
+    // 1. Get media files for the project
+    let media_files =
+        queries::get_media_files_for_project(&conn, &project_id).map_err(|e| e.to_string())?;
+
+    let media_file = match media_files.first() {
+        Some(mf) => mf,
+        None => return Ok(None),
+    };
+
+    // 2. Get speakers for the project and build id -> label map
+    let speakers =
+        queries::get_speakers_for_project(&conn, &project_id).map_err(|e| e.to_string())?;
+    let speaker_label_map: std::collections::HashMap<String, String> = speakers
+        .iter()
+        .map(|s| (s.id.clone(), s.label.clone()))
+        .collect();
+
+    // 3. Get segments for the media file
+    let db_segments =
+        queries::get_segments_for_media(&conn, &media_file.id).map_err(|e| e.to_string())?;
+
+    // 4. Build output segments with nested words
+    let mut segment_outputs = Vec::with_capacity(db_segments.len());
+    for seg in &db_segments {
+        let words = queries::get_words_for_segment(&conn, &seg.id).map_err(|e| e.to_string())?;
+        let word_outputs: Vec<WordOutput> = words
+            .into_iter()
+            .map(|w| WordOutput {
+                word: w.word,
+                start_ms: w.start_ms,
+                end_ms: w.end_ms,
+                confidence: w.confidence,
+            })
+            .collect();
+
+        let speaker_label = seg
+            .speaker_id
+            .as_ref()
+            .and_then(|sid| speaker_label_map.get(sid))
+            .cloned();
+
+        segment_outputs.push(SegmentOutput {
+            id: seg.id.clone(),
+            text: seg.text.clone(),
+            start_ms: seg.start_ms,
+            end_ms: seg.end_ms,
+            speaker: speaker_label,
+            words: word_outputs,
+        });
+    }
+
+    // 5. Build speaker outputs
+    let speaker_outputs: Vec<SpeakerOutput> = speakers
+        .into_iter()
+        .map(|s| SpeakerOutput {
+            id: s.id,
+            label: s.label,
+            display_name: s.display_name,
+            color: s.color,
+        })
+        .collect();
+
+    Ok(Some(ProjectTranscript {
+        file_path: media_file.file_path.clone(),
+        segments: segment_outputs,
+        speakers: speaker_outputs,
+    }))
+}
+
+// ── File-based project commands ─────────────────────────────────
+
+#[tauri::command]
+pub fn save_project_file(path: String, project: ProjectFile) -> Result<(), String> {
+    let json = serde_json::to_string_pretty(&project).map_err(|e| e.to_string())?;
+    fs::write(&path, json).map_err(|e| format!("Failed to save project: {e}"))
+}
+
+#[tauri::command]
+pub fn load_project_file(path: String) -> Result<ProjectFile, String> {
+    let json = fs::read_to_string(&path).map_err(|e| format!("Failed to read project: {e}"))?;
+    serde_json::from_str(&json).map_err(|e| format!("Failed to parse project: {e}"))
 }

src-tauri/src/commands/system.rs

@@ -22,9 +22,7 @@ pub fn llama_start(
     threads: Option<u32>,
 ) -> Result<LlamaStatus, String> {
     let config = LlamaConfig {
-        binary_path: PathBuf::from(
-            binary_path.unwrap_or_else(|| "llama-server".to_string()),
-        ),
+        binary_path: PathBuf::from(binary_path.unwrap_or_else(|| "llama-server".to_string())),
         model_path: PathBuf::from(model_path),
         port: port.unwrap_or(0),
         n_gpu_layers: n_gpu_layers.unwrap_or(0),

src-tauri/src/commands/transcribe.rs

@@ -1,12 +1,10 @@
 use serde_json::{json, Value};
+use tauri::{AppHandle, Emitter};
 
 use crate::sidecar::messages::IPCMessage;
-use crate::sidecar::SidecarManager;
+use crate::sidecar::sidecar;
 
 /// Start transcription of an audio file via the Python sidecar.
-///
-/// This is a blocking command — it starts the sidecar if needed,
-/// sends the transcribe request, and waits for the result.
 #[tauri::command]
 pub fn transcribe_file(
     file_path: String,
@@ -14,17 +12,8 @@ pub fn transcribe_file(
     device: Option<String>,
     language: Option<String>,
 ) -> Result<Value, String> {
-    // Determine Python sidecar path (relative to app)
-    let python_path = std::env::current_dir()
-        .map_err(|e| e.to_string())?
-        .join("../python")
-        .canonicalize()
-        .map_err(|e| format!("Cannot find python directory: {e}"))?;
-
-    let python_path_str = python_path.to_string_lossy().to_string();
-
-    let manager = SidecarManager::new();
-    manager.start(&python_path_str)?;
+    let manager = sidecar();
+    manager.ensure_running()?;
 
     let request_id = uuid::Uuid::new_v4().to_string();
     let msg = IPCMessage::new(
@@ -44,16 +33,48 @@ pub fn transcribe_file(
     if response.msg_type == "error" {
         return Err(format!(
             "Transcription error: {}",
-            response.payload.get("message").and_then(|v| v.as_str()).unwrap_or("unknown")
+            response
+                .payload
+                .get("message")
+                .and_then(|v| v.as_str())
+                .unwrap_or("unknown")
         ));
     }
 
     Ok(response.payload)
 }
 
+/// Download and validate the diarization model via the Python sidecar.
+#[tauri::command]
+pub fn download_diarize_model(hf_token: String) -> Result<Value, String> {
+    let manager = sidecar();
+    manager.ensure_running()?;
+
+    let request_id = uuid::Uuid::new_v4().to_string();
+    let msg = IPCMessage::new(
+        &request_id,
+        "diarize.download",
+        json!({
+            "hf_token": hf_token,
+        }),
+    );
+
+    let response = manager.send_and_receive(&msg)?;
+
+    if response.msg_type == "error" {
+        return Ok(json!({
+            "ok": false,
+            "error": response.payload.get("message").and_then(|v| v.as_str()).unwrap_or("unknown"),
+        }));
+    }
+
+    Ok(json!({ "ok": true }))
+}
+
 /// Run the full transcription + diarization pipeline via the Python sidecar.
 #[tauri::command]
-pub fn run_pipeline(
+pub async fn run_pipeline(
+    app: AppHandle,
     file_path: String,
     model: Option<String>,
     device: Option<String>,
@@ -62,17 +83,10 @@ pub fn run_pipeline(
     min_speakers: Option<u32>,
     max_speakers: Option<u32>,
     skip_diarization: Option<bool>,
+    hf_token: Option<String>,
 ) -> Result<Value, String> {
-    let python_path = std::env::current_dir()
-        .map_err(|e| e.to_string())?
-        .join("../python")
-        .canonicalize()
-        .map_err(|e| format!("Cannot find python directory: {e}"))?;
-
-    let python_path_str = python_path.to_string_lossy().to_string();
-
-    let manager = SidecarManager::new();
-    manager.start(&python_path_str)?;
+    let manager = sidecar();
+    manager.ensure_running()?;
 
     let request_id = uuid::Uuid::new_v4().to_string();
     let msg = IPCMessage::new(
@@ -88,17 +102,38 @@ pub fn run_pipeline(
             "min_speakers": min_speakers,
             "max_speakers": max_speakers,
             "skip_diarization": skip_diarization.unwrap_or(false),
+            "hf_token": hf_token,
         }),
     );
 
-    let response = manager.send_and_receive(&msg)?;
+    // Run the blocking sidecar I/O on a separate thread so the async runtime
+    // can deliver emitted events to the webview while processing is ongoing.
+    let app_handle = app.clone();
+    tauri::async_runtime::spawn_blocking(move || {
+        let response = manager.send_and_receive_with_progress(&msg, |msg| {
+            let event_name = match msg.msg_type.as_str() {
+                "pipeline.segment" => "pipeline-segment",
+                "pipeline.speaker_update" => "pipeline-speaker-update",
+                _ => "pipeline-progress",
+            };
+            if let Err(e) = app_handle.emit(event_name, &msg.payload) {
+                eprintln!("[sidecar-rs] Failed to emit {event_name}: {e}");
+            }
+        })?;
 
-    if response.msg_type == "error" {
-        return Err(format!(
-            "Pipeline error: {}",
-            response.payload.get("message").and_then(|v| v.as_str()).unwrap_or("unknown")
-        ));
-    }
+        if response.msg_type == "error" {
+            return Err(format!(
+                "Pipeline error: {}",
+                response
+                    .payload
+                    .get("message")
+                    .and_then(|v| v.as_str())
+                    .unwrap_or("unknown")
+            ));
+        }
 
-    Ok(response.payload)
+        Ok(response.payload)
+    })
+    .await
+    .map_err(|e| format!("Pipeline task failed: {e}"))?
 }

src-tauri/src/db/queries.rs

@@ -85,6 +85,57 @@ pub fn delete_project(conn: &Connection, id: &str) -> Result<(), DatabaseError>
     Ok(())
 }
 
+// ── Media Files ──────────────────────────────────────────────────
+
+pub fn create_media_file(
+    conn: &Connection,
+    project_id: &str,
+    file_path: &str,
+) -> Result<MediaFile, DatabaseError> {
+    let id = Uuid::new_v4().to_string();
+    let now = Utc::now().to_rfc3339();
+    conn.execute(
+        "INSERT INTO media_files (id, project_id, file_path, created_at) VALUES (?1, ?2, ?3, ?4)",
+        params![id, project_id, file_path, now],
+    )?;
+    Ok(MediaFile {
+        id,
+        project_id: project_id.to_string(),
+        file_path: file_path.to_string(),
+        file_hash: None,
+        duration_ms: None,
+        sample_rate: None,
+        channels: None,
+        format: None,
+        file_size: None,
+        created_at: now,
+    })
+}
+
+pub fn get_media_files_for_project(
+    conn: &Connection,
+    project_id: &str,
+) -> Result<Vec<MediaFile>, DatabaseError> {
+    let mut stmt = conn.prepare(
+        "SELECT id, project_id, file_path, file_hash, duration_ms, sample_rate, channels, format, file_size, created_at FROM media_files WHERE project_id = ?1 ORDER BY created_at",
+    )?;
+    let rows = stmt.query_map(params![project_id], |row| {
+        Ok(MediaFile {
+            id: row.get(0)?,
+            project_id: row.get(1)?,
+            file_path: row.get(2)?,
+            file_hash: row.get(3)?,
+            duration_ms: row.get(4)?,
+            sample_rate: row.get(5)?,
+            channels: row.get(6)?,
+            format: row.get(7)?,
+            file_size: row.get(8)?,
+            created_at: row.get(9)?,
+        })
+    })?;
+    Ok(rows.collect::<Result<Vec<_>, _>>()?)
+}
+
 // ── Speakers ──────────────────────────────────────────────────────
 
 pub fn create_speaker(
@@ -194,6 +245,39 @@ pub fn reassign_speaker(
     Ok(())
 }
 
+// ── Segments (create) ────────────────────────────────────────────
+
+pub fn create_segment(
+    conn: &Connection,
+    project_id: &str,
+    media_file_id: &str,
+    speaker_id: Option<&str>,
+    start_ms: i64,
+    end_ms: i64,
+    text: &str,
+    segment_index: i32,
+) -> Result<Segment, DatabaseError> {
+    let id = Uuid::new_v4().to_string();
+    conn.execute(
+        "INSERT INTO segments (id, project_id, media_file_id, speaker_id, start_ms, end_ms, text, is_edited, segment_index) VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7, 0, ?8)",
+        params![id, project_id, media_file_id, speaker_id, start_ms, end_ms, text, segment_index],
+    )?;
+    Ok(Segment {
+        id,
+        project_id: project_id.to_string(),
+        media_file_id: media_file_id.to_string(),
+        speaker_id: speaker_id.map(String::from),
+        start_ms,
+        end_ms,
+        text: text.to_string(),
+        original_text: None,
+        confidence: None,
+        is_edited: false,
+        edited_at: None,
+        segment_index,
+    })
+}
+
 // ── Words ─────────────────────────────────────────────────────────
 
 pub fn get_words_for_segment(
@@ -217,6 +301,31 @@ pub fn get_words_for_segment(
     Ok(rows.collect::<Result<Vec<_>, _>>()?)
 }
 
+pub fn create_word(
+    conn: &Connection,
+    segment_id: &str,
+    word: &str,
+    start_ms: i64,
+    end_ms: i64,
+    confidence: Option<f64>,
+    word_index: i32,
+) -> Result<Word, DatabaseError> {
+    let id = Uuid::new_v4().to_string();
+    conn.execute(
+        "INSERT INTO words (id, segment_id, word, start_ms, end_ms, confidence, word_index) VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7)",
+        params![id, segment_id, word, start_ms, end_ms, confidence, word_index],
+    )?;
+    Ok(Word {
+        id,
+        segment_id: segment_id.to_string(),
+        word: word.to_string(),
+        start_ms,
+        end_ms,
+        confidence,
+        word_index,
+    })
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

src-tauri/src/db/schema.rs

@@ -96,11 +96,7 @@ pub fn create_tables(conn: &Connection) -> Result<(), DatabaseError> {
     )?;
 
     // Initialize schema version if empty
-    let count: i32 = conn.query_row(
-        "SELECT COUNT(*) FROM schema_version",
-        [],
-        |row| row.get(0),
-    )?;
+    let count: i32 = conn.query_row("SELECT COUNT(*) FROM schema_version", [], |row| row.get(0))?;
     if count == 0 {
         conn.execute(
             "INSERT INTO schema_version (version) VALUES (?1)",

src-tauri/src/lib.rs

@@ -4,24 +4,56 @@ pub mod llama;
 pub mod sidecar;
 pub mod state;
 
+use tauri::window::Color;
+use tauri::Manager;
+
 use commands::ai::{ai_chat, ai_configure, ai_list_providers};
 use commands::export::export_transcript;
-use commands::project::{create_project, get_project, list_projects};
+use commands::project::{
+    create_project, delete_project, get_project, list_projects, load_project_file,
+    load_project_transcript, save_project_file, save_project_transcript, update_segment,
+};
 use commands::settings::{load_settings, save_settings};
 use commands::system::{get_data_dir, llama_list_models, llama_start, llama_status, llama_stop};
-use commands::transcribe::{run_pipeline, transcribe_file};
+use commands::transcribe::{download_diarize_model, run_pipeline, transcribe_file};
+use state::AppState;
 
 #[cfg_attr(mobile, tauri::mobile_entry_point)]
 pub fn run() {
+    let app_state = AppState::new().expect("Failed to initialize app state");
+
     tauri::Builder::default()
         .plugin(tauri_plugin_opener::init())
         .plugin(tauri_plugin_dialog::init())
+        .manage(app_state)
+        .setup(|app| {
+            // Tell the sidecar manager where Tauri placed bundled resources
+            // and where to extract the sidecar archive
+            if let (Ok(resource_dir), Ok(data_dir)) =
+                (app.path().resource_dir(), app.path().app_local_data_dir())
+            {
+                sidecar::init_dirs(resource_dir, data_dir);
+            }
+
+            // Set the webview background to match the app's dark theme
+            if let Some(window) = app.get_webview_window("main") {
+                let _ = window.set_background_color(Some(Color(10, 10, 35, 255)));
+            }
+            Ok(())
+        })
         .invoke_handler(tauri::generate_handler![
             create_project,
             get_project,
             list_projects,
+            delete_project,
+            save_project_transcript,
+            load_project_transcript,
+            update_segment,
+            save_project_file,
+            load_project_file,
             transcribe_file,
             run_pipeline,
+            download_diarize_model,
             export_transcript,
             ai_chat,
             ai_list_providers,

src-tauri/src/llama/mod.rs

@@ -237,11 +237,7 @@ impl LlamaManager {
 
     /// Get the current status.
     pub fn status(&self) -> LlamaStatus {
-        let running = self
-            .process
-            .lock()
-            .ok()
-            .map_or(false, |p| p.is_some());
+        let running = self.process.lock().ok().map_or(false, |p| p.is_some());
         let port = self.port.lock().ok().map_or(0, |p| *p);
         let model = self
             .model_path

src-tauri/src/sidecar/mod.rs

@@ -2,76 +2,346 @@ pub mod ipc;
 pub mod messages;
 
 use std::io::{BufRead, BufReader, Write};
-use std::process::{Child, Command, Stdio};
-use std::sync::Mutex;
+use std::path::{Path, PathBuf};
+use std::process::{Child, ChildStdin, Command, Stdio};
+use std::sync::{Mutex, OnceLock};
 
 use crate::sidecar::messages::IPCMessage;
 
-/// Manages the Python sidecar process lifecycle.
+/// Resource directory set by the Tauri app during setup.
+static RESOURCE_DIR: OnceLock<PathBuf> = OnceLock::new();
+/// App data directory for extracting the sidecar archive.
+static DATA_DIR: OnceLock<PathBuf> = OnceLock::new();
+
+/// Initialize directories for sidecar resolution.
+/// Must be called from the Tauri setup before any sidecar operations.
+pub fn init_dirs(resource_dir: PathBuf, data_dir: PathBuf) {
+    RESOURCE_DIR.set(resource_dir).ok();
+    DATA_DIR.set(data_dir).ok();
+}
+
+/// Get the global sidecar manager singleton.
+pub fn sidecar() -> &'static SidecarManager {
+    static INSTANCE: OnceLock<SidecarManager> = OnceLock::new();
+    INSTANCE.get_or_init(SidecarManager::new)
+}
+
+/// Manages the sidecar process lifecycle.
+///
+/// Supports two modes:
+/// - **Production**: spawns a frozen PyInstaller binary (no Python required)
+/// - **Dev mode**: spawns system Python with `-m voice_to_notes.main`
+///
+/// Dev mode is active when compiled in debug mode or when `VOICE_TO_NOTES_DEV=1`.
 pub struct SidecarManager {
     process: Mutex<Option<Child>>,
+    stdin: Mutex<Option<ChildStdin>>,
+    reader: Mutex<Option<BufReader<std::process::ChildStdout>>>,
 }
 
 impl SidecarManager {
     pub fn new() -> Self {
         Self {
             process: Mutex::new(None),
+            stdin: Mutex::new(None),
+            reader: Mutex::new(None),
         }
     }
 
-    /// Spawn the Python sidecar process.
-    pub fn start(&self, python_path: &str) -> Result<(), String> {
-        let child = Command::new("python3")
-            .arg("-m")
-            .arg("voice_to_notes.main")
-            .current_dir(python_path)
-            .env("PYTHONPATH", python_path)
+    /// Check if we should use dev mode (system Python).
+    fn is_dev_mode() -> bool {
+        cfg!(debug_assertions) || std::env::var("VOICE_TO_NOTES_DEV").is_ok()
+    }
+
+    /// Resolve the frozen sidecar binary path (production mode).
+    ///
+    /// First checks if the sidecar is already extracted to the app data directory.
+    /// If not, looks for `sidecar.zip` in the Tauri resource directory and extracts it.
+    fn resolve_sidecar_path() -> Result<PathBuf, String> {
+        let binary_name = if cfg!(target_os = "windows") {
+            "voice-to-notes-sidecar.exe"
+        } else {
+            "voice-to-notes-sidecar"
+        };
+
+        // Versioned extraction directory prevents stale sidecar after app updates
+        let extract_dir = DATA_DIR
+            .get()
+            .ok_or("App data directory not initialized")?
+            .join(format!("sidecar-{}", env!("CARGO_PKG_VERSION")));
+
+        let binary_path = extract_dir.join(binary_name);
+
+        // Already extracted — use it directly
+        if binary_path.exists() {
+            return Ok(binary_path);
+        }
+
+        // Find sidecar.zip in resource dir or next to exe
+        let zip_path = Self::find_sidecar_zip()?;
+        Self::extract_zip(&zip_path, &extract_dir)?;
+
+        if !binary_path.exists() {
+            return Err(format!(
+                "Sidecar binary not found after extraction at {}",
+                binary_path.display()
+            ));
+        }
+
+        // Make executable on Unix
+        #[cfg(unix)]
+        {
+            use std::os::unix::fs::PermissionsExt;
+            if let Ok(meta) = std::fs::metadata(&binary_path) {
+                let mut perms = meta.permissions();
+                perms.set_mode(0o755);
+                let _ = std::fs::set_permissions(&binary_path, perms);
+            }
+        }
+
+        Ok(binary_path)
+    }
+
+    /// Locate the bundled sidecar.zip archive.
+    fn find_sidecar_zip() -> Result<PathBuf, String> {
+        let mut candidates: Vec<PathBuf> = Vec::new();
+
+        if let Some(resource_dir) = RESOURCE_DIR.get() {
+            candidates.push(resource_dir.join("sidecar.zip"));
+        }
+        if let Ok(exe) = std::env::current_exe() {
+            if let Some(exe_dir) = exe.parent() {
+                candidates.push(exe_dir.join("sidecar.zip"));
+            }
+        }
+
+        for path in &candidates {
+            if path.exists() {
+                return Ok(path.clone());
+            }
+        }
+
+        Err(format!(
+            "Sidecar archive not found. Checked:\n{}",
+            candidates
+                .iter()
+                .map(|p| format!("  {}", p.display()))
+                .collect::<Vec<_>>()
+                .join("\n"),
+        ))
+    }
+
+    /// Extract a zip archive to the given directory.
+    fn extract_zip(zip_path: &Path, dest: &Path) -> Result<(), String> {
+        eprintln!(
+            "[sidecar-rs] Extracting sidecar from {} to {}",
+            zip_path.display(),
+            dest.display()
+        );
+
+        // Clean destination so we don't mix old and new files
+        if dest.exists() {
+            std::fs::remove_dir_all(dest)
+                .map_err(|e| format!("Failed to clean extraction dir: {e}"))?;
+        }
+        std::fs::create_dir_all(dest)
+            .map_err(|e| format!("Failed to create extraction dir: {e}"))?;
+
+        let file =
+            std::fs::File::open(zip_path).map_err(|e| format!("Cannot open sidecar zip: {e}"))?;
+        let mut archive =
+            zip::ZipArchive::new(file).map_err(|e| format!("Invalid sidecar zip: {e}"))?;
+
+        for i in 0..archive.len() {
+            let mut entry = archive
+                .by_index(i)
+                .map_err(|e| format!("Zip entry error: {e}"))?;
+
+            let name = entry.name().to_string();
+            let outpath = dest.join(&name);
+
+            if entry.is_dir() {
+                std::fs::create_dir_all(&outpath)
+                    .map_err(|e| format!("Cannot create dir {}: {e}", outpath.display()))?;
+            } else {
+                if let Some(parent) = outpath.parent() {
+                    std::fs::create_dir_all(parent)
+                        .map_err(|e| format!("Cannot create dir {}: {e}", parent.display()))?;
+                }
+                let mut outfile = std::fs::File::create(&outpath)
+                    .map_err(|e| format!("Cannot create {}: {e}", outpath.display()))?;
+                std::io::copy(&mut entry, &mut outfile)
+                    .map_err(|e| format!("Write error for {}: {e}", name))?;
+            }
+        }
+
+        eprintln!("[sidecar-rs] Sidecar extracted successfully");
+        Ok(())
+    }
+
+    /// Find a working Python command for the current platform.
+    fn find_python_command() -> &'static str {
+        if cfg!(target_os = "windows") {
+            "python"
+        } else {
+            "python3"
+        }
+    }
+
+    /// Resolve the Python sidecar directory for dev mode.
+    fn resolve_python_dir() -> Result<std::path::PathBuf, String> {
+        let manifest_dir = env!("CARGO_MANIFEST_DIR");
+        let python_dir = std::path::Path::new(manifest_dir)
+            .join("../python")
+            .canonicalize()
+            .map_err(|e| format!("Cannot find python directory: {e}"))?;
+
+        if python_dir.exists() {
+            return Ok(python_dir);
+        }
+
+        // Fallback: relative to current exe
+        let exe = std::env::current_exe().map_err(|e| e.to_string())?;
+        let alt = exe
+            .parent()
+            .ok_or_else(|| "No parent dir".to_string())?
+            .join("../python")
+            .canonicalize()
+            .map_err(|e| format!("Cannot find python directory: {e}"))?;
+
+        Ok(alt)
+    }
+
+    /// Ensure the sidecar is running, starting it if needed.
+    pub fn ensure_running(&self) -> Result<(), String> {
+        if self.is_running() {
+            return Ok(());
+        }
+
+        if Self::is_dev_mode() {
+            self.start_python_dev()
+        } else {
+            let path = Self::resolve_sidecar_path()?;
+            self.start_binary(&path)
+        }
+    }
+
+    /// Spawn the frozen sidecar binary (production mode).
+    fn start_binary(&self, path: &std::path::Path) -> Result<(), String> {
+        self.stop().ok();
+        eprintln!("[sidecar-rs] Starting frozen sidecar: {}", path.display());
+
+        let child = Command::new(path)
             .stdin(Stdio::piped())
             .stdout(Stdio::piped())
-            .stderr(Stdio::inherit()) // Let sidecar logs go to parent's stderr
+            .stderr(Stdio::inherit())
             .spawn()
-            .map_err(|e| format!("Failed to start sidecar: {e}"))?;
+            .map_err(|e| format!("Failed to start sidecar binary: {e}"))?;
 
-        let mut proc = self.process.lock().map_err(|e| e.to_string())?;
-        *proc = Some(child);
+        self.attach(child)?;
+        self.wait_for_ready()
+    }
 
-        // Wait for the "ready" message
-        self.wait_for_ready()?;
+    /// Spawn the Python sidecar in dev mode (system Python).
+    fn start_python_dev(&self) -> Result<(), String> {
+        self.stop().ok();
+        let python_dir = Self::resolve_python_dir()?;
+        let python_cmd = Self::find_python_command();
+        eprintln!(
+            "[sidecar-rs] Starting dev sidecar: {} -m voice_to_notes.main ({})",
+            python_cmd,
+            python_dir.display()
+        );
 
+        let child = Command::new(python_cmd)
+            .arg("-m")
+            .arg("voice_to_notes.main")
+            .current_dir(&python_dir)
+            .env("PYTHONPATH", &python_dir)
+            .stdin(Stdio::piped())
+            .stdout(Stdio::piped())
+            .stderr(Stdio::inherit())
+            .spawn()
+            .map_err(|e| format!("Failed to start Python sidecar: {e}"))?;
+
+        self.attach(child)?;
+        self.wait_for_ready()
+    }
+
+    /// Take ownership of a spawned child's stdin/stdout and store the process handle.
+    fn attach(&self, mut child: Child) -> Result<(), String> {
+        let stdin = child.stdin.take().ok_or("Failed to get sidecar stdin")?;
+        let stdout = child.stdout.take().ok_or("Failed to get sidecar stdout")?;
+        let buf_reader = BufReader::new(stdout);
+
+        {
+            let mut proc = self.process.lock().map_err(|e| e.to_string())?;
+            *proc = Some(child);
+        }
+        {
+            let mut s = self.stdin.lock().map_err(|e| e.to_string())?;
+            *s = Some(stdin);
+        }
+        {
+            let mut r = self.reader.lock().map_err(|e| e.to_string())?;
+            *r = Some(buf_reader);
+        }
         Ok(())
     }
 
     /// Wait for the sidecar to send its ready message.
     fn wait_for_ready(&self) -> Result<(), String> {
-        let mut proc = self.process.lock().map_err(|e| e.to_string())?;
-        if let Some(ref mut child) = *proc {
-            if let Some(ref mut stdout) = child.stdout {
-                let reader = BufReader::new(stdout);
-                for line in reader.lines() {
-                    let line = line.map_err(|e| format!("Read error: {e}"))?;
-                    if line.is_empty() {
-                        continue;
-                    }
-                    if let Ok(msg) = serde_json::from_str::<IPCMessage>(&line) {
-                        if msg.msg_type == "ready" {
-                            return Ok(());
-                        }
-                    }
-                    // If we got a non-ready message, something's wrong but don't block forever
-                    break;
+        let mut reader_guard = self.reader.lock().map_err(|e| e.to_string())?;
+        if let Some(ref mut reader) = *reader_guard {
+            let mut line = String::new();
+            loop {
+                line.clear();
+                let bytes = reader
+                    .read_line(&mut line)
+                    .map_err(|e| format!("Read error: {e}"))?;
+                if bytes == 0 {
+                    return Err("Sidecar closed stdout before sending ready".to_string());
                 }
+                let trimmed = line.trim();
+                if trimmed.is_empty() {
+                    continue;
+                }
+                if let Ok(msg) = serde_json::from_str::<IPCMessage>(trimmed) {
+                    if msg.msg_type == "ready" {
+                        return Ok(());
+                    }
+                }
+                // Non-JSON or non-ready line — skip and keep waiting
+                eprintln!(
+                    "[sidecar-rs] Skipping pre-ready line: {}",
+                    &trimmed[..trimmed.len().min(200)]
+                );
+                continue;
             }
         }
         Err("Sidecar did not send ready message".to_string())
     }
 
     /// Send a message to the sidecar and read the response.
-    /// This is a blocking call.
+    /// This is a blocking call. Progress messages are skipped.
     pub fn send_and_receive(&self, msg: &IPCMessage) -> Result<IPCMessage, String> {
-        let mut proc = self.process.lock().map_err(|e| e.to_string())?;
-        if let Some(ref mut child) = *proc {
-            // Write message to stdin
-            if let Some(ref mut stdin) = child.stdin {
+        self.send_and_receive_with_progress(msg, |_| {})
+    }
+
+    /// Send a message and receive the response, calling a callback for intermediate messages.
+    /// Intermediate messages include progress, pipeline.segment, and pipeline.speaker_update.
+    pub fn send_and_receive_with_progress<F>(
+        &self,
+        msg: &IPCMessage,
+        on_intermediate: F,
+    ) -> Result<IPCMessage, String>
+    where
+        F: Fn(&IPCMessage),
+    {
+        // Write to stdin
+        {
+            let mut stdin_guard = self.stdin.lock().map_err(|e| e.to_string())?;
+            if let Some(ref mut stdin) = *stdin_guard {
                 let json = serde_json::to_string(msg).map_err(|e| e.to_string())?;
                 stdin
                     .write_all(json.as_bytes())
@@ -83,12 +353,13 @@ impl SidecarManager {
             } else {
                 return Err("Sidecar stdin not available".to_string());
             }
+        }
 
-            // Read response from stdout
-            if let Some(ref mut stdout) = child.stdout {
-                let mut reader = BufReader::new(stdout);
+        // Read from stdout
+        {
+            let mut reader_guard = self.reader.lock().map_err(|e| e.to_string())?;
+            if let Some(ref mut reader) = *reader_guard {
                 let mut line = String::new();
-                // Read lines until we get a response (skip progress messages, collect them)
                 loop {
                     line.clear();
                     let bytes_read = reader
@@ -104,37 +375,48 @@ impl SidecarManager {
                     let response: IPCMessage =
                         serde_json::from_str(trimmed).map_err(|e| format!("Parse error: {e}"))?;
 
-                    // If it's a progress message, we could emit it as an event
-                    // For now, skip progress and return the final result/error
-                    if response.msg_type != "progress" {
+                    // Forward intermediate messages via callback, return the final result/error
+                    let is_intermediate = matches!(
+                        response.msg_type.as_str(),
+                        "progress" | "pipeline.segment" | "pipeline.speaker_update"
+                    );
+                    if is_intermediate {
+                        on_intermediate(&response);
+                    } else {
                         return Ok(response);
                     }
                 }
             } else {
-                return Err("Sidecar stdout not available".to_string());
+                Err("Sidecar stdout not available".to_string())
             }
-        } else {
-            Err("Sidecar not running".to_string())
         }
     }
 
     /// Stop the sidecar process.
     pub fn stop(&self) -> Result<(), String> {
-        let mut proc = self.process.lock().map_err(|e| e.to_string())?;
-        if let Some(ref mut child) = proc.take() {
-            // Close stdin to signal EOF
-            drop(child.stdin.take());
-            // Wait briefly for clean exit, then kill
-            match child.wait() {
-                Ok(_) => Ok(()),
-                Err(e) => {
-                    let _ = child.kill();
-                    Err(format!("Sidecar did not exit cleanly: {e}"))
+        // Drop stdin to signal EOF
+        {
+            let mut stdin_guard = self.stdin.lock().map_err(|e| e.to_string())?;
+            *stdin_guard = None;
+        }
+        // Drop reader
+        {
+            let mut reader_guard = self.reader.lock().map_err(|e| e.to_string())?;
+            *reader_guard = None;
+        }
+        // Wait for process to exit
+        {
+            let mut proc = self.process.lock().map_err(|e| e.to_string())?;
+            if let Some(ref mut child) = proc.take() {
+                match child.wait() {
+                    Ok(_) => {}
+                    Err(_) => {
+                        let _ = child.kill();
+                    }
                 }
             }
-        } else {
-            Ok(())
         }
+        Ok(())
     }
 
     pub fn is_running(&self) -> bool {

src-tauri/src/state.rs

@@ -3,17 +3,26 @@ use std::sync::Mutex;
 
 use rusqlite::Connection;
 
+use crate::db;
+use crate::llama::LlamaManager;
+
 /// Shared application state managed by Tauri.
 pub struct AppState {
-    pub db: Mutex<Option<Connection>>,
+    pub db: Mutex<Connection>,
     pub data_dir: PathBuf,
 }
 
 impl AppState {
-    pub fn new(data_dir: PathBuf) -> Self {
-        Self {
-            db: Mutex::new(None),
+    pub fn new() -> Result<Self, String> {
+        let data_dir = LlamaManager::data_dir();
+        std::fs::create_dir_all(&data_dir).map_err(|e| format!("Cannot create data dir: {e}"))?;
+
+        let db_path = data_dir.join("voice_to_notes.db");
+        let conn = db::open_database(&db_path).map_err(|e| format!("Cannot open database: {e}"))?;
+
+        Ok(Self {
+            db: Mutex::new(conn),
             data_dir,
-        }
+        })
     }
 }

src-tauri/tauri.conf.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://schema.tauri.app/config/2",
   "productName": "Voice to Notes",
-  "version": "0.1.0",
+  "version": "0.2.0",
   "identifier": "com.voicetonotes.app",
   "build": {
     "beforeDevCommand": "npm run dev",
@@ -16,7 +16,9 @@
         "width": 1200,
         "height": 800,
         "minWidth": 800,
-        "minHeight": 600
+        "minHeight": 600,
+        "decorations": true,
+        "transparent": false
       }
     ],
     "security": {
@@ -29,7 +31,7 @@
   },
   "bundle": {
     "active": true,
-    "targets": "all",
+    "targets": ["deb", "nsis", "msi", "dmg"],
     "icon": [
       "icons/32x32.png",
       "icons/128x128.png",
@@ -40,14 +42,12 @@
     "category": "Utility",
     "shortDescription": "Transcribe audio/video with speaker identification",
     "longDescription": "Voice to Notes is a desktop application that transcribes audio and video recordings with speaker identification, synchronized playback, and AI-powered analysis. Export to SRT, WebVTT, ASS captions, or plain text.",
+    "resources": ["sidecar.zip"],
     "copyright": "Voice to Notes Contributors",
     "license": "MIT",
     "linux": {
       "deb": {
-        "depends": ["python3", "python3-pip"]
-      },
-      "appimage": {
-        "bundleMediaFramework": true
+        "depends": []
       }
     },
     "windows": {

src/app.html

@@ -1,5 +1,5 @@
 <!doctype html>
-<html lang="en">
+<html lang="en" style="margin:0;padding:0;background:#0a0a23;height:100%;">
   <head>
     <meta charset="utf-8" />
     <link rel="icon" href="%sveltekit.assets%/favicon.png" />
@@ -7,7 +7,7 @@
     <title>Voice to Notes</title>
     %sveltekit.head%
   </head>
-  <body data-sveltekit-preload-data="hover">
+  <body data-sveltekit-preload-data="hover" style="margin:0;padding:0;background:#0a0a23;overflow:hidden;">
     <div style="display: contents">%sveltekit.body%</div>
   </body>
 </html>

src/lib/components/AIChatPanel.svelte

@@ -1,6 +1,7 @@
 <script lang="ts">
   import { invoke } from '@tauri-apps/api/core';
   import { segments, speakers } from '$lib/stores/transcript';
+  import { settings } from '$lib/stores/settings';
 
   interface ChatMessage {
     role: 'user' | 'assistant';
@@ -43,9 +44,23 @@
         content: m.content,
       }));
 
+      // Ensure the provider is configured with current credentials before chatting
+      const s = $settings;
+      const configMap: Record<string, Record<string, string>> = {
+        openai: { api_key: s.openai_api_key, model: s.openai_model },
+        anthropic: { api_key: s.anthropic_api_key, model: s.anthropic_model },
+        litellm: { api_key: s.litellm_api_key, api_base: s.litellm_api_base, model: s.litellm_model },
+        local: { model: s.local_model_path, base_url: 'http://localhost:8080' },
+      };
+      const config = configMap[s.ai_provider];
+      if (config) {
+        await invoke('ai_configure', { provider: s.ai_provider, config });
+      }
+
       const result = await invoke<{ response: string }>('ai_chat', {
         messages: chatMessages,
         transcriptContext: getTranscriptContext(),
+        provider: s.ai_provider,
       });
 
       messages = [...messages, { role: 'assistant', content: result.response }];
@@ -73,6 +88,88 @@
     messages = [];
   }
 
+  function formatMarkdown(text: string): string {
+    // Split into lines for block-level processing
+    const lines = text.split('\n');
+    const result: string[] = [];
+    let inList = false;
+
+    for (let i = 0; i < lines.length; i++) {
+      let line = lines[i];
+
+      // Headers
+      if (line.startsWith('### ')) {
+        if (inList) { result.push('</ul>'); inList = false; }
+        const content = applyInlineFormatting(line.slice(4));
+        result.push(`<h4>${content}</h4>`);
+        continue;
+      }
+      if (line.startsWith('## ')) {
+        if (inList) { result.push('</ul>'); inList = false; }
+        const content = applyInlineFormatting(line.slice(3));
+        result.push(`<h3>${content}</h3>`);
+        continue;
+      }
+      if (line.startsWith('# ')) {
+        if (inList) { result.push('</ul>'); inList = false; }
+        const content = applyInlineFormatting(line.slice(2));
+        result.push(`<h2>${content}</h2>`);
+        continue;
+      }
+
+      // List items (- or *)
+      if (/^[\-\*] /.test(line)) {
+        if (!inList) { result.push('<ul>'); inList = true; }
+        const content = applyInlineFormatting(line.slice(2));
+        result.push(`<li>${content}</li>`);
+        continue;
+      }
+
+      // Numbered list items
+      if (/^\d+\.\s/.test(line)) {
+        if (!inList) { result.push('<ol>'); inList = true; }
+        const content = applyInlineFormatting(line.replace(/^\d+\.\s/, ''));
+        result.push(`<li>${content}</li>`);
+        continue;
+      }
+
+      // Non-list line: close any open list
+      if (inList) {
+        // Check if previous list was ordered or unordered
+        const lastOpen = result.findLast(r => r === '<ul>' || r === '<ol>');
+        result.push(lastOpen === '<ol>' ? '</ol>' : '</ul>');
+        inList = false;
+      }
+
+      // Empty line = paragraph break
+      if (line.trim() === '') {
+        result.push('<br>');
+        continue;
+      }
+
+      // Regular text line
+      result.push(applyInlineFormatting(line));
+    }
+
+    // Close any trailing open list
+    if (inList) {
+      const lastOpen = result.findLast(r => r === '<ul>' || r === '<ol>');
+      result.push(lastOpen === '<ol>' ? '</ol>' : '</ul>');
+    }
+
+    return result.join('\n');
+  }
+
+  function applyInlineFormatting(text: string): string {
+    // Code blocks (backtick) — process first to avoid conflicts
+    text = text.replace(/`([^`]+)`/g, '<code>$1</code>');
+    // Bold (**text**)
+    text = text.replace(/\*\*([^*]+)\*\*/g, '<strong>$1</strong>');
+    // Italic (*text*) — only single asterisks not already consumed by bold
+    text = text.replace(/\*([^*]+)\*/g, '<em>$1</em>');
+    return text;
+  }
+
   // Quick action buttons
   async function summarize() {
     inputText = 'Please summarize this transcript in bullet points.';
@@ -107,7 +204,11 @@
     {:else}
       {#each messages as msg}
         <div class="message {msg.role}">
-          <div class="message-content">{msg.content}</div>
+          {#if msg.role === 'assistant'}
+            <div class="message-content">{@html formatMarkdown(msg.content)}</div>
+          {:else}
+            <div class="message-content">{msg.content}</div>
+          {/if}
         </div>
       {/each}
       {#if isLoading}
@@ -177,47 +278,101 @@
   }
   .empty-state {
     text-align: center;
-    color: #666;
-    font-size: 0.8rem;
-    padding: 1rem 0;
+    color: #888;
+    font-size: 0.85rem;
+    padding: 2rem 1rem;
+  }
+  .empty-state p {
+    margin-bottom: 1rem;
   }
   .quick-actions {
     display: flex;
-    gap: 0.5rem;
+    gap: 0.75rem;
     justify-content: center;
-    margin-top: 0.5rem;
+    margin-top: 1rem;
   }
   .quick-btn {
     background: rgba(233, 69, 96, 0.15);
     border: 1px solid rgba(233, 69, 96, 0.3);
     color: #e94560;
-    padding: 0.3rem 0.6rem;
-    border-radius: 4px;
+    padding: 0.45rem 0.85rem;
+    border-radius: 6px;
     cursor: pointer;
-    font-size: 0.75rem;
+    font-size: 0.8rem;
+    transition: background 0.15s;
   }
   .quick-btn:hover {
     background: rgba(233, 69, 96, 0.25);
   }
   .message {
-    margin-bottom: 0.5rem;
-    padding: 0.5rem 0.75rem;
-    border-radius: 6px;
+    margin-bottom: 0.75rem;
+    padding: 0.75rem 1rem;
+    border-radius: 8px;
     font-size: 0.8rem;
-    line-height: 1.4;
+    line-height: 1.55;
   }
   .message.user {
     background: rgba(233, 69, 96, 0.15);
-    margin-left: 1rem;
+    border-left: 3px solid rgba(233, 69, 96, 0.4);
   }
   .message.assistant {
     background: rgba(255, 255, 255, 0.05);
-    margin-right: 1rem;
+    border-left: 3px solid rgba(255, 255, 255, 0.1);
   }
   .message.loading {
     opacity: 0.6;
     font-style: italic;
   }
+
+  /* Markdown styles inside assistant messages */
+  .message.assistant :global(h2) {
+    font-size: 1rem;
+    font-weight: 600;
+    margin: 0.6rem 0 0.3rem;
+    color: #f0f0f0;
+  }
+  .message.assistant :global(h3) {
+    font-size: 0.9rem;
+    font-weight: 600;
+    margin: 0.5rem 0 0.25rem;
+    color: #e8e8e8;
+  }
+  .message.assistant :global(h4) {
+    font-size: 0.85rem;
+    font-weight: 600;
+    margin: 0.4rem 0 0.2rem;
+    color: #e0e0e0;
+  }
+  .message.assistant :global(strong) {
+    color: #f0f0f0;
+    font-weight: 600;
+  }
+  .message.assistant :global(em) {
+    color: #ccc;
+    font-style: italic;
+  }
+  .message.assistant :global(code) {
+    background: rgba(0, 0, 0, 0.3);
+    color: #e94560;
+    padding: 0.1rem 0.35rem;
+    border-radius: 3px;
+    font-size: 0.75rem;
+    font-family: 'Fira Code', 'Cascadia Code', 'Consolas', monospace;
+  }
+  .message.assistant :global(ul),
+  .message.assistant :global(ol) {
+    margin: 0.35rem 0;
+    padding-left: 1.3rem;
+  }
+  .message.assistant :global(li) {
+    margin-bottom: 0.25rem;
+    line-height: 1.5;
+  }
+  .message.assistant :global(br) {
+    display: block;
+    content: '';
+    margin-top: 0.35rem;
+  }
   .chat-input {
     display: flex;
     gap: 0.5rem;

src/lib/components/ProgressOverlay.svelte

@@ -7,16 +7,88 @@
   }
 
   let { visible = false, percent = 0, stage = '', message = '' }: Props = $props();
+
+  // Pipeline steps in order
+  const pipelineSteps = [
+    { key: 'loading_model', label: 'Load transcription model' },
+    { key: 'transcribing', label: 'Transcribe audio' },
+    { key: 'loading_diarization', label: 'Load speaker detection model' },
+    { key: 'diarizing', label: 'Identify speakers' },
+    { key: 'merging', label: 'Merge results' },
+  ];
+
+  const stepOrder = pipelineSteps.map(s => s.key);
+
+  // Track the highest step index we've reached (never goes backward)
+  let highestStepIdx = $state(-1);
+
+  // Map non-step stages to step indices for progress tracking
+  function stageToStepIdx(s: string): number {
+    const direct = stepOrder.indexOf(s);
+    if (direct >= 0) return direct;
+    // 'pipeline' stage appears before known steps — don't change highwater mark
+    return -1;
+  }
+
+  $effect(() => {
+    if (!visible) {
+      highestStepIdx = -1;
+      return;
+    }
+    const idx = stageToStepIdx(stage);
+    if (idx > highestStepIdx) {
+      highestStepIdx = idx;
+    }
+  });
+
+  function getStepStatus(stepIdx: number): 'pending' | 'active' | 'done' {
+    if (stepIdx < highestStepIdx) return 'done';
+    if (stepIdx === highestStepIdx) return 'active';
+    return 'pending';
+  }
+
+  // User-friendly display of current stage
+  const stageLabels: Record<string, string> = {
+    'pipeline': 'Initializing...',
+    'loading_model': 'Loading Model',
+    'transcribing': 'Transcribing',
+    'loading_diarization': 'Loading Diarization',
+    'diarizing': 'Speaker Detection',
+    'merging': 'Merging Results',
+    'done': 'Complete',
+  };
+
+  let displayStage = $derived(stageLabels[stage] || stage || 'Processing...');
 </script>
 
 {#if visible}
   <div class="overlay">
     <div class="progress-card">
-      <h3>{stage}</h3>
-      <div class="bar-track">
-        <div class="bar-fill" style="width: {percent}%"></div>
+      <div class="spinner-row">
+        <div class="spinner"></div>
+        <h3>{displayStage}</h3>
       </div>
-      <p>{percent}% — {message}</p>
+
+      <div class="steps">
+        {#each pipelineSteps as step, idx}
+          {@const status = getStepStatus(idx)}
+          <div class="step" class:step-done={status === 'done'} class:step-active={status === 'active'}>
+            <span class="step-icon">
+              {#if status === 'done'}
+                ✓
+              {:else if status === 'active'}
+                ⟳
+              {:else}
+                ·
+              {/if}
+            </span>
+            <span class="step-label">{step.label}</span>
+          </div>
+        {/each}
+      </div>
+
+      <p class="status-text">{message || 'Please wait...'}</p>
+      <p class="hint-text">This may take several minutes for large files</p>
     </div>
   </div>
 {/if}
@@ -25,34 +97,81 @@
   .overlay {
     position: fixed;
     inset: 0;
-    background: rgba(0, 0, 0, 0.7);
+    background: rgba(0, 0, 0, 0.8);
     display: flex;
     align-items: center;
     justify-content: center;
-    z-index: 1000;
+    z-index: 9999;
   }
   .progress-card {
     background: #16213e;
-    padding: 2rem;
+    padding: 2rem 2.5rem;
     border-radius: 12px;
-    min-width: 400px;
+    min-width: 380px;
+    max-width: 440px;
     color: #e0e0e0;
+    border: 1px solid #2a3a5e;
+    box-shadow: 0 8px 32px rgba(0, 0, 0, 0.5);
   }
-  h3 { margin: 0 0 1rem; text-transform: capitalize; }
-  .bar-track {
-    height: 8px;
-    background: #0f3460;
-    border-radius: 4px;
-    overflow: hidden;
+  .spinner-row {
+    display: flex;
+    align-items: center;
+    gap: 0.75rem;
+    margin-bottom: 1.25rem;
   }
-  .bar-fill {
-    height: 100%;
-    background: #e94560;
-    transition: width 0.3s;
+  .spinner {
+    width: 20px;
+    height: 20px;
+    border: 3px solid #2a3a5e;
+    border-top-color: #e94560;
+    border-radius: 50%;
+    animation: spin 0.8s linear infinite;
+    flex-shrink: 0;
   }
-  p {
+  @keyframes spin {
+    to { transform: rotate(360deg); }
+  }
+  h3 {
+    margin: 0;
+    font-size: 1.1rem;
+  }
+  .steps {
+    display: flex;
+    flex-direction: column;
+    gap: 0.4rem;
+    margin-bottom: 1rem;
+  }
+  .step {
+    display: flex;
+    align-items: center;
+    gap: 0.5rem;
+    font-size: 0.85rem;
+    color: #555;
+  }
+  .step-done {
+    color: #4ecdc4;
+  }
+  .step-active {
+    color: #e0e0e0;
+    font-weight: 500;
+  }
+  .step-icon {
+    width: 1.2rem;
+    text-align: center;
+    flex-shrink: 0;
+  }
+  .step-active .step-icon {
+    animation: spin 1.5s linear infinite;
+    display: inline-block;
+  }
+  .status-text {
+    margin: 0.75rem 0 0;
+    font-size: 0.85rem;
+    color: #b0b0b0;
+  }
+  .hint-text {
     margin: 0.5rem 0 0;
-    font-size: 0.875rem;
-    color: #999;
+    font-size: 0.75rem;
+    color: #555;
   }
 </style>

src/lib/components/SettingsModal.svelte

@@ -1,4 +1,6 @@
 <script lang="ts">
+  import { invoke } from '@tauri-apps/api/core';
+  import { openUrl } from '@tauri-apps/plugin-opener';
   import { settings, saveSettings, type AppSettings } from '$lib/stores/settings';
 
   interface Props {
@@ -9,7 +11,34 @@
   let { visible, onClose }: Props = $props();
 
   let localSettings = $state<AppSettings>({ ...$settings });
-  let activeTab = $state<'transcription' | 'ai' | 'local'>('transcription');
+  let activeTab = $state<'transcription' | 'speakers' | 'ai' | 'local'>('transcription');
+  let modelStatus = $state<'idle' | 'downloading' | 'success' | 'error'>('idle');
+  let modelError = $state('');
+  let revealedFields = $state<Set<string>>(new Set());
+
+  async function testAndDownloadModel() {
+    if (!localSettings.hf_token) {
+      modelStatus = 'error';
+      modelError = 'Please enter a HuggingFace token first.';
+      return;
+    }
+    modelStatus = 'downloading';
+    modelError = '';
+    try {
+      const result = await invoke<{ ok: boolean; error?: string }>('download_diarize_model', {
+        hfToken: localSettings.hf_token,
+      });
+      if (result.ok) {
+        modelStatus = 'success';
+      } else {
+        modelStatus = 'error';
+        modelError = result.error || 'Unknown error';
+      }
+    } catch (err) {
+      modelStatus = 'error';
+      modelError = String(err);
+    }
+  }
 
   // Sync when settings store changes
   $effect(() => {
@@ -46,6 +75,9 @@
         <button class="tab" class:active={activeTab === 'transcription'} onclick={() => activeTab = 'transcription'}>
           Transcription
         </button>
+        <button class="tab" class:active={activeTab === 'speakers'} onclick={() => activeTab = 'speakers'}>
+          Speakers
+        </button>
         <button class="tab" class:active={activeTab === 'ai'} onclick={() => activeTab = 'ai'}>
           AI Provider
         </button>
@@ -77,10 +109,72 @@
             <label for="stt-lang">Language (blank = auto-detect)</label>
             <input id="stt-lang" type="text" bind:value={localSettings.transcription_language} placeholder="e.g., en, es, fr" />
           </div>
-          <div class="field checkbox">
+        {:else if activeTab === 'speakers'}
+          <div class="field">
+            <label for="hf-token">HuggingFace Token</label>
+            <div class="input-reveal">
+              <input id="hf-token" type={revealedFields.has('hf-token') ? 'text' : 'password'} bind:value={localSettings.hf_token} placeholder="hf_..." />
+              <button type="button" class="reveal-btn" onclick={() => { const s = new Set(revealedFields); s.has('hf-token') ? s.delete('hf-token') : s.add('hf-token'); revealedFields = s; }}>{revealedFields.has('hf-token') ? 'Hide' : 'Show'}</button>
+            </div>
+          </div>
+          <div class="info-box">
+            <p class="info-title">Setup (one-time)</p>
+            <p>Speaker detection uses <strong>pyannote.audio</strong> models hosted on HuggingFace. You must accept the license for each model:</p>
+            <ol>
+              <li>Create a free account at <!-- svelte-ignore a11y_no_static_element_interactions --><a class="ext-link" onclick={() => openUrl('https://huggingface.co/join')}>huggingface.co</a></li>
+              <li>Accept the license on <strong>all three</strong> of these pages:
+                <ul>
+                  <!-- svelte-ignore a11y_no_static_element_interactions -->
+                  <li><a class="ext-link" onclick={() => openUrl('https://huggingface.co/pyannote/speaker-diarization-3.1')}>pyannote/speaker-diarization-3.1</a></li>
+                  <!-- svelte-ignore a11y_no_static_element_interactions -->
+                  <li><a class="ext-link" onclick={() => openUrl('https://huggingface.co/pyannote/segmentation-3.0')}>pyannote/segmentation-3.0</a></li>
+                  <!-- svelte-ignore a11y_no_static_element_interactions -->
+                  <li><a class="ext-link" onclick={() => openUrl('https://huggingface.co/pyannote/speaker-diarization-community-1')}>pyannote/speaker-diarization-community-1</a></li>
+                </ul>
+              </li>
+              <!-- svelte-ignore a11y_no_static_element_interactions -->
+              <li>Create a token at <a class="ext-link" onclick={() => openUrl('https://huggingface.co/settings/tokens')}>huggingface.co/settings/tokens</a> (read access)</li>
+              <li>Paste the token above and click <strong>Test & Download</strong></li>
+            </ol>
+          </div>
+          <button
+            class="btn-download"
+            onclick={testAndDownloadModel}
+            disabled={modelStatus === 'downloading'}
+          >
+            {#if modelStatus === 'downloading'}
+              Downloading model...
+            {:else}
+              Test & Download Model
+            {/if}
+          </button>
+          {#if modelStatus === 'success'}
+            <p class="status-success">Model downloaded successfully. Speaker detection is ready.</p>
+          {/if}
+          {#if modelStatus === 'error'}
+            <p class="status-error">{modelError}</p>
+          {/if}
+          <div class="field" style="margin-top: 1rem;">
+            <label for="num-speakers">Number of speakers</label>
+            <select
+              id="num-speakers"
+              value={localSettings.num_speakers === null || localSettings.num_speakers === 0 ? '0' : String(localSettings.num_speakers)}
+              onchange={(e) => {
+                const v = parseInt((e.target as HTMLSelectElement).value, 10);
+                localSettings.num_speakers = v === 0 ? null : v;
+              }}
+            >
+              <option value="0">Auto-detect</option>
+              {#each Array.from({ length: 20 }, (_, i) => i + 1) as n}
+                <option value={String(n)}>{n}</option>
+              {/each}
+            </select>
+            <p class="hint">Hint the expected number of speakers to speed up diarization clustering.</p>
+          </div>
+          <div class="field checkbox" style="margin-top: 1rem;">
             <label>
               <input type="checkbox" bind:checked={localSettings.skip_diarization} />
-              Skip speaker diarization (faster, no speaker labels)
+              Skip speaker detection (faster, no speaker labels)
             </label>
           </div>
         {:else if activeTab === 'ai'}
@@ -90,14 +184,17 @@
               <option value="local">Local (llama-server)</option>
               <option value="openai">OpenAI</option>
               <option value="anthropic">Anthropic</option>
-              <option value="litellm">LiteLLM</option>
+              <option value="litellm">OpenAI Compatible</option>
             </select>
           </div>
 
           {#if localSettings.ai_provider === 'openai'}
             <div class="field">
               <label for="openai-key">OpenAI API Key</label>
-              <input id="openai-key" type="password" bind:value={localSettings.openai_api_key} placeholder="sk-..." />
+              <div class="input-reveal">
+                <input id="openai-key" type={revealedFields.has('openai-key') ? 'text' : 'password'} bind:value={localSettings.openai_api_key} placeholder="sk-..." />
+                <button type="button" class="reveal-btn" onclick={() => { const s = new Set(revealedFields); s.has('openai-key') ? s.delete('openai-key') : s.add('openai-key'); revealedFields = s; }}>{revealedFields.has('openai-key') ? 'Hide' : 'Show'}</button>
+              </div>
             </div>
             <div class="field">
               <label for="openai-model">Model</label>
@@ -106,13 +203,27 @@
           {:else if localSettings.ai_provider === 'anthropic'}
             <div class="field">
               <label for="anthropic-key">Anthropic API Key</label>
-              <input id="anthropic-key" type="password" bind:value={localSettings.anthropic_api_key} placeholder="sk-ant-..." />
+              <div class="input-reveal">
+                <input id="anthropic-key" type={revealedFields.has('anthropic-key') ? 'text' : 'password'} bind:value={localSettings.anthropic_api_key} placeholder="sk-ant-..." />
+                <button type="button" class="reveal-btn" onclick={() => { const s = new Set(revealedFields); s.has('anthropic-key') ? s.delete('anthropic-key') : s.add('anthropic-key'); revealedFields = s; }}>{revealedFields.has('anthropic-key') ? 'Hide' : 'Show'}</button>
+              </div>
             </div>
             <div class="field">
               <label for="anthropic-model">Model</label>
               <input id="anthropic-model" type="text" bind:value={localSettings.anthropic_model} />
             </div>
           {:else if localSettings.ai_provider === 'litellm'}
+            <div class="field">
+              <label for="litellm-base">API Base URL</label>
+              <input id="litellm-base" type="text" bind:value={localSettings.litellm_api_base} placeholder="https://your-litellm-proxy.example.com" />
+            </div>
+            <div class="field">
+              <label for="litellm-key">API Key</label>
+              <div class="input-reveal">
+                <input id="litellm-key" type={revealedFields.has('litellm-key') ? 'text' : 'password'} bind:value={localSettings.litellm_api_key} placeholder="sk-..." />
+                <button type="button" class="reveal-btn" onclick={() => { const s = new Set(revealedFields); s.has('litellm-key') ? s.delete('litellm-key') : s.add('litellm-key'); revealedFields = s; }}>{revealedFields.has('litellm-key') ? 'Hide' : 'Show'}</button>
+              </div>
+            </div>
             <div class="field">
               <label for="litellm-model">Model</label>
               <input id="litellm-model" type="text" bind:value={localSettings.litellm_model} placeholder="provider/model-name" />
@@ -220,11 +331,36 @@
     color: #aaa;
     margin-bottom: 0.3rem;
   }
+  .input-reveal {
+    display: flex;
+    gap: 0;
+  }
+  .input-reveal input {
+    flex: 1;
+    border-top-right-radius: 0;
+    border-bottom-right-radius: 0;
+  }
+  .reveal-btn {
+    background: #0f3460;
+    border: 1px solid #4a5568;
+    border-left: none;
+    color: #aaa;
+    padding: 0.5rem 0.6rem;
+    border-radius: 0 4px 4px 0;
+    cursor: pointer;
+    font-size: 0.75rem;
+    white-space: nowrap;
+  }
+  .reveal-btn:hover {
+    color: #e0e0e0;
+    background: #1a4a7a;
+  }
   .field input,
   .field select {
     width: 100%;
     background: #1a1a2e;
     color: #e0e0e0;
+    color-scheme: dark;
     border: 1px solid #4a5568;
     border-radius: 4px;
     padding: 0.5rem;
@@ -252,6 +388,79 @@
     color: #666;
     line-height: 1.4;
   }
+  .info-box {
+    background: rgba(233, 69, 96, 0.05);
+    border: 1px solid #2a3a5e;
+    border-radius: 6px;
+    padding: 0.75rem 1rem;
+    margin-bottom: 1rem;
+    font-size: 0.8rem;
+    color: #b0b0b0;
+    line-height: 1.5;
+  }
+  .info-box p {
+    margin: 0 0 0.5rem;
+  }
+  .info-box p:last-child {
+    margin-bottom: 0;
+  }
+  .info-box .info-title {
+    color: #e0e0e0;
+    font-weight: 600;
+    font-size: 0.8rem;
+  }
+  .info-box ol {
+    margin: 0.25rem 0 0.5rem;
+    padding-left: 1.25rem;
+  }
+  .info-box li {
+    margin-bottom: 0.25rem;
+  }
+  .info-box strong {
+    color: #e0e0e0;
+  }
+  .ext-link {
+    color: #e94560;
+    cursor: pointer;
+    text-decoration: underline;
+  }
+  .ext-link:hover {
+    color: #ff6b81;
+  }
+  .info-box ul {
+    margin: 0.25rem 0;
+    padding-left: 1.25rem;
+  }
+  .btn-download {
+    background: #0f3460;
+    border: 1px solid #4a5568;
+    color: #e0e0e0;
+    padding: 0.5rem 1rem;
+    border-radius: 6px;
+    cursor: pointer;
+    font-size: 0.85rem;
+    width: 100%;
+    margin-bottom: 0.5rem;
+  }
+  .btn-download:hover:not(:disabled) {
+    background: #1a4a7a;
+    border-color: #e94560;
+  }
+  .btn-download:disabled {
+    opacity: 0.6;
+    cursor: not-allowed;
+  }
+  .status-success {
+    color: #4ecdc4;
+    font-size: 0.8rem;
+    margin: 0.25rem 0;
+  }
+  .status-error {
+    color: #e94560;
+    font-size: 0.8rem;
+    margin: 0.25rem 0;
+    word-break: break-word;
+  }
   .modal-footer {
     display: flex;
     justify-content: flex-end;

src/lib/components/SpeakerManager.svelte

@@ -1,5 +1,6 @@
 <script lang="ts">
   import { speakers } from '$lib/stores/transcript';
+  import { settings } from '$lib/stores/settings';
   import type { Speaker } from '$lib/types/transcript';
 
   let editingSpeakerId = $state<string | null>(null);
@@ -34,7 +35,14 @@
 <div class="speaker-manager">
   <h3>Speakers</h3>
   {#if $speakers.length === 0}
-    <p class="empty-hint">No speakers detected yet</p>
+    <p class="empty-hint">No speakers detected</p>
+    {#if $settings.skip_diarization}
+      <p class="setup-hint">Speaker detection is disabled. Enable it in Settings &gt; Speakers.</p>
+    {:else if !$settings.hf_token}
+      <p class="setup-hint">Speaker detection requires a HuggingFace token. Configure it in Settings &gt; Speakers.</p>
+    {:else}
+      <p class="setup-hint">Speaker detection ran but found no distinct speakers, or the model may need to be downloaded. Check Settings &gt; Speakers.</p>
+    {/if}
   {:else}
     <ul class="speaker-list">
       {#each $speakers as speaker (speaker.id)}
@@ -78,6 +86,19 @@
   .empty-hint {
     color: #666;
     font-size: 0.875rem;
+    margin-bottom: 0.25rem;
+  }
+  .setup-hint {
+    color: #555;
+    font-size: 0.75rem;
+    line-height: 1.4;
+  }
+  .setup-hint code {
+    background: rgba(233, 69, 96, 0.15);
+    color: #e94560;
+    padding: 0.1rem 0.3rem;
+    border-radius: 3px;
+    font-size: 0.7rem;
   }
   .speaker-list {
     list-style: none;

src/lib/components/TranscriptEditor.svelte

@@ -60,12 +60,14 @@
   function finishEditing(segmentId: string) {
     const trimmed = editText.trim();
     if (trimmed) {
-      // Update the segment text in the store
       segments.update(segs => segs.map(s => {
         if (s.id !== segmentId) return s;
+        const newWordTexts = trimmed.split(/\s+/);
+        const newWords = redistributeWords(s, newWordTexts);
         return {
           ...s,
           text: trimmed,
+          words: newWords,
           original_text: s.original_text ?? s.text,
           is_edited: true,
           edited_at: new Date().toISOString(),
@@ -76,6 +78,106 @@
     editingSegmentId = null;
   }
 
+  /**
+   * Redistribute word timing after an edit.
+   *
+   * Uses a diff-like alignment between old and new word lists:
+   * - Unchanged words keep their original timing
+   * - Spelling fixes (same position, same count) keep timing
+   * - Split words (1 old → N new) divide the original time range proportionally
+   * - Inserted words with no match get interpolated timing
+   */
+  function redistributeWords(segment: Segment, newWordTexts: string[]): Word[] {
+    const oldWords = segment.words;
+
+    // Same word count — preserve per-word timing (spelling fixes)
+    if (newWordTexts.length === oldWords.length) {
+      return oldWords.map((w, i) => ({ ...w, word: newWordTexts[i] }));
+    }
+
+    // Align old words to new words using a simple greedy match.
+    // Build a mapping: for each old word, which new words does it cover?
+    const oldTexts = oldWords.map(w => w.word.toLowerCase());
+    const newTexts = newWordTexts.map(w => w.toLowerCase());
+
+    // Walk both lists, greedily matching old words to new words
+    const result: Word[] = [];
+    let oldIdx = 0;
+    let newIdx = 0;
+
+    while (newIdx < newTexts.length) {
+      if (oldIdx < oldTexts.length && oldTexts[oldIdx] === newTexts[newIdx]) {
+        // Exact match — keep original timing
+        result.push({ ...oldWords[oldIdx], word: newWordTexts[newIdx], word_index: newIdx });
+        oldIdx++;
+        newIdx++;
+      } else if (oldIdx < oldTexts.length) {
+        // Check if old word was split into multiple new words.
+        // E.g., "gonna" → "going to": see if concatenating upcoming new words
+        // matches the old word (or close enough — just check if old word's chars
+        // are consumed by the next few new words).
+        let splitCount = 0;
+        let combined = '';
+        for (let k = newIdx; k < newTexts.length && k - newIdx < 5; k++) {
+          combined += (k > newIdx ? '' : '') + newTexts[k];
+          if (combined.length >= oldTexts[oldIdx].length) {
+            splitCount = k - newIdx + 1;
+            break;
+          }
+        }
+
+        if (splitCount > 1) {
+          // Split: distribute the old word's time range proportionally
+          const ow = oldWords[oldIdx];
+          const totalDuration = ow.end_ms - ow.start_ms;
+          for (let k = 0; k < splitCount; k++) {
+            const fraction = 1 / splitCount;
+            result.push({
+              id: `${segment.id}-word-${newIdx + k}`,
+              segment_id: segment.id,
+              word: newWordTexts[newIdx + k],
+              start_ms: Math.round(ow.start_ms + totalDuration * fraction * k),
+              end_ms: Math.round(ow.start_ms + totalDuration * fraction * (k + 1)),
+              confidence: ow.confidence,
+              word_index: newIdx + k,
+            });
+          }
+          oldIdx++;
+          newIdx += splitCount;
+        } else {
+          // No match found — interpolate timing from neighbors
+          const prevEnd = result.length > 0 ? result[result.length - 1].end_ms : segment.start_ms;
+          const nextStart = oldIdx < oldWords.length ? oldWords[oldIdx].start_ms : segment.end_ms;
+          result.push({
+            id: `${segment.id}-word-${newIdx}`,
+            segment_id: segment.id,
+            word: newWordTexts[newIdx],
+            start_ms: prevEnd,
+            end_ms: nextStart,
+            confidence: 1.0,
+            word_index: newIdx,
+          });
+          newIdx++;
+        }
+      } else {
+        // No more old words — use end of segment
+        const prevEnd = result.length > 0 ? result[result.length - 1].end_ms : segment.start_ms;
+        result.push({
+          id: `${segment.id}-word-${newIdx}`,
+          segment_id: segment.id,
+          word: newWordTexts[newIdx],
+          start_ms: prevEnd,
+          end_ms: segment.end_ms,
+          confidence: 1.0,
+          word_index: newIdx,
+        });
+        newIdx++;
+      }
+    }
+
+    return result;
+  }
+
   function handleEditKeydown(e: KeyboardEvent, segmentId: string) {
     if (e.key === 'Escape') {
       editingSegmentId = null;
@@ -217,6 +319,8 @@
   .segment-text {
     line-height: 1.6;
     padding-left: 0.75rem;
+    word-wrap: break-word;
+    overflow-wrap: break-word;
   }
   .word {
     cursor: pointer;

src/lib/components/WaveformPlayer.svelte

@@ -12,6 +12,8 @@
 
   let container: HTMLDivElement;
   let wavesurfer: WaveSurfer | null = $state(null);
+  let isReady = $state(false);
+  let isLoading = $state(false);
   let currentTime = $state('0:00');
   let totalTime = $state('0:00');
 
@@ -31,6 +33,7 @@
       barWidth: 2,
       barGap: 1,
       barRadius: 2,
+      backend: 'WebAudio',
     });
 
     wavesurfer.on('timeupdate', (time: number) => {
@@ -39,6 +42,8 @@
     });
 
     wavesurfer.on('ready', () => {
+      isReady = true;
+      isLoading = false;
       const dur = wavesurfer!.getDuration();
       durationMs.set(Math.round(dur * 1000));
       totalTime = formatTime(dur);
@@ -48,8 +53,12 @@
     wavesurfer.on('pause', () => isPlaying.set(false));
     wavesurfer.on('finish', () => isPlaying.set(false));
 
+    wavesurfer.on('loading', () => {
+      isReady = false;
+    });
+
     if (audioUrl) {
-      wavesurfer.load(audioUrl);
+      loadAudio(audioUrl);
     }
   });
 
@@ -57,20 +66,21 @@
     wavesurfer?.destroy();
   });
 
-  /** Toggle play/pause. Exposed for keyboard shortcuts. */
+  /** Toggle play/pause from current position. Exposed for keyboard shortcuts. */
   export function togglePlayPause() {
-    wavesurfer?.playPause();
+    if (!wavesurfer || !isReady) return;
+    wavesurfer.playPause();
   }
 
   function skipBack() {
-    if (wavesurfer) {
+    if (wavesurfer && isReady) {
       const time = Math.max(0, wavesurfer.getCurrentTime() - 5);
       wavesurfer.setTime(time);
     }
   }
 
   function skipForward() {
-    if (wavesurfer) {
+    if (wavesurfer && isReady) {
       const time = Math.min(wavesurfer.getDuration(), wavesurfer.getCurrentTime() + 5);
       wavesurfer.setTime(time);
     }
@@ -78,16 +88,17 @@
 
   /** Seek to a specific time in milliseconds. Called from transcript click-to-seek. */
   export function seekTo(timeMs: number) {
-    if (wavesurfer) {
-      wavesurfer.setTime(timeMs / 1000);
-      if (!wavesurfer.isPlaying()) {
-        wavesurfer.play();
-      }
+    if (!wavesurfer || !isReady) {
+      console.warn('[voice-to-notes] seekTo ignored — audio not ready yet');
+      return;
     }
+    wavesurfer.setTime(timeMs / 1000);
   }
 
   /** Load a new audio file. */
   export function loadAudio(url: string) {
+    isReady = false;
+    isLoading = true;
     wavesurfer?.load(url);
   }
 </script>
@@ -95,11 +106,17 @@
 <div class="waveform-player">
   <div class="waveform-container" bind:this={container}></div>
   <div class="controls">
-    <button class="control-btn" onclick={skipBack} title="Back 5s">⏪</button>
-    <button class="control-btn play-btn" onclick={togglePlayPause} title="Play/Pause">
-      {#if $isPlaying}⏸{:else}▶{/if}
+    <button class="control-btn" onclick={skipBack} title="Back 5s" disabled={!isReady}>⏪</button>
+    <button class="control-btn play-btn" onclick={togglePlayPause} title="Play/Pause" disabled={!isReady}>
+      {#if !isReady}
+        ⏳
+      {:else if $isPlaying}
+        ⏸
+      {:else}
+        ▶
+      {/if}
     </button>
-    <button class="control-btn" onclick={skipForward} title="Forward 5s">⏩</button>
+    <button class="control-btn" onclick={skipForward} title="Forward 5s" disabled={!isReady}>⏩</button>
     <span class="time">{currentTime} / {totalTime}</span>
   </div>
 </div>
@@ -129,9 +146,13 @@
     cursor: pointer;
     font-size: 1rem;
   }
-  .control-btn:hover {
+  .control-btn:hover:not(:disabled) {
     background: #1a4a7a;
   }
+  .control-btn:disabled {
+    opacity: 0.4;
+    cursor: not-allowed;
+  }
   .play-btn {
     padding: 0.4rem 1rem;
     font-size: 1.2rem;

src/lib/stores/settings.ts

@@ -8,12 +8,16 @@ export interface AppSettings {
   openai_model: string;
   anthropic_model: string;
   litellm_model: string;
+  litellm_api_key: string;
+  litellm_api_base: string;
   local_model_path: string;
   local_binary_path: string;
   transcription_model: string;
   transcription_device: string;
   transcription_language: string;
   skip_diarization: boolean;
+  hf_token: string;
+  num_speakers: number | null;
 }
 
 const defaults: AppSettings = {
@@ -23,12 +27,16 @@ const defaults: AppSettings = {
   openai_model: 'gpt-4o-mini',
   anthropic_model: 'claude-sonnet-4-6',
   litellm_model: 'gpt-4o-mini',
+  litellm_api_key: '',
+  litellm_api_base: '',
   local_model_path: '',
   local_binary_path: 'llama-server',
   transcription_model: 'base',
   transcription_device: 'cpu',
   transcription_language: '',
   skip_diarization: false,
+  hf_token: '',
+  num_speakers: null,
 };
 
 export const settings = writable<AppSettings>({ ...defaults });
@@ -45,4 +53,20 @@ export async function loadSettings(): Promise<void> {
 export async function saveSettings(s: AppSettings): Promise<void> {
   settings.set(s);
   await invoke('save_settings', { settings: s });
+
+  // Configure the AI provider in the Python sidecar
+  const configMap: Record<string, Record<string, string>> = {
+    openai: { api_key: s.openai_api_key, model: s.openai_model },
+    anthropic: { api_key: s.anthropic_api_key, model: s.anthropic_model },
+    litellm: { api_key: s.litellm_api_key, api_base: s.litellm_api_base, model: s.litellm_model },
+    local: { model: s.local_model_path, base_url: 'http://localhost:8080' },
+  };
+  const config = configMap[s.ai_provider];
+  if (config) {
+    try {
+      await invoke('ai_configure', { provider: s.ai_provider, config });
+    } catch {
+      // Sidecar may not be running yet — provider will be configured on first use
+    }
+  }
 }

src/routes/+layout.svelte

@@ -0,0 +1,40 @@
+<script>
+  let { children } = $props();
+</script>
+
+{@render children()}
+
+<style>
+  :global(html, body) {
+    margin: 0;
+    padding: 0;
+    background: #0a0a23;
+    color: #e0e0e0;
+    color-scheme: dark;
+    font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, Oxygen,
+      Ubuntu, Cantarell, sans-serif;
+    overflow: hidden;
+    height: 100%;
+  }
+
+  :global(*, *::before, *::after) {
+    box-sizing: border-box;
+  }
+
+  :global(::-webkit-scrollbar) {
+    width: 8px;
+  }
+
+  :global(::-webkit-scrollbar-track) {
+    background: #0a0a23;
+  }
+
+  :global(::-webkit-scrollbar-thumb) {
+    background: #2a3a5e;
+    border-radius: 4px;
+  }
+
+  :global(::-webkit-scrollbar-thumb:hover) {
+    background: #4a5568;
+  }
+</style>

src/routes/+page.svelte

@@ -1,5 +1,6 @@
 <script lang="ts">
-  import { invoke } from '@tauri-apps/api/core';
+  import { invoke, convertFileSrc } from '@tauri-apps/api/core';
+  import { listen } from '@tauri-apps/api/event';
   import { open, save } from '@tauri-apps/plugin-dialog';
   import WaveformPlayer from '$lib/components/WaveformPlayer.svelte';
   import TranscriptEditor from '$lib/components/TranscriptEditor.svelte';
@@ -10,12 +11,18 @@
   import { segments, speakers } from '$lib/stores/transcript';
   import { settings, loadSettings } from '$lib/stores/settings';
   import type { Segment, Speaker } from '$lib/types/transcript';
-  import { onMount } from 'svelte';
+  import { onMount, tick } from 'svelte';
 
+  let appReady = $state(false);
   let waveformPlayer: WaveformPlayer;
   let audioUrl = $state('');
   let showSettings = $state(false);
 
+  // Project management state
+  let currentProjectPath = $state<string | null>(null);
+  let currentProjectName = $state('');
+  let audioFilePath = $state('');
+
   onMount(() => {
     loadSettings();
 
@@ -42,8 +49,8 @@
 
     // Close export dropdown on outside click
     function handleClickOutside(e: MouseEvent) {
+      const target = e.target as HTMLElement;
       if (showExportMenu) {
-        const target = e.target as HTMLElement;
         if (!target.closest('.export-dropdown')) {
           showExportMenu = false;
         }
@@ -53,6 +60,8 @@
     document.addEventListener('keydown', handleKeyDown);
     document.addEventListener('click', handleClickOutside);
 
+    appReady = true;
+
     return () => {
       document.removeEventListener('keydown', handleKeyDown);
       document.removeEventListener('click', handleClickOutside);
@@ -66,10 +75,136 @@
   // Speaker color palette for auto-assignment
   const speakerColors = ['#e94560', '#4ecdc4', '#ffe66d', '#a8e6cf', '#ff8b94', '#c7ceea', '#ffd93d', '#6bcb77'];
 
+  async function saveProject() {
+    const defaultName = currentProjectName || 'Untitled';
+    const outputPath = await save({
+      defaultPath: `${defaultName}.vtn`,
+      filters: [{ name: 'Voice to Notes Project', extensions: ['vtn'] }],
+    });
+    if (!outputPath) return;
+
+    const projectData = {
+      version: 1,
+      name: outputPath.split(/[\\/]/).pop()?.replace('.vtn', '') || defaultName,
+      audio_file: audioFilePath,
+      created_at: new Date().toISOString(),
+      segments: $segments.map(seg => {
+        const speaker = $speakers.find(s => s.id === seg.speaker_id);
+        return {
+          text: seg.text,
+          start_ms: seg.start_ms,
+          end_ms: seg.end_ms,
+          speaker: speaker?.label ?? null,
+          is_edited: seg.is_edited,
+          words: seg.words.map(w => ({
+            word: w.word,
+            start_ms: w.start_ms,
+            end_ms: w.end_ms,
+            confidence: w.confidence ?? 0,
+          })),
+        };
+      }),
+      speakers: $speakers.map(s => ({
+        label: s.label,
+        display_name: s.display_name,
+        color: s.color || '#e94560',
+      })),
+    };
+
+    try {
+      await invoke('save_project_file', { path: outputPath, project: projectData });
+      currentProjectPath = outputPath;
+      currentProjectName = projectData.name;
+    } catch (err) {
+      console.error('Failed to save project:', err);
+      alert(`Failed to save: ${err}`);
+    }
+  }
+
+  async function openProject() {
+    const filePath = await open({
+      filters: [{ name: 'Voice to Notes Project', extensions: ['vtn'] }],
+      multiple: false,
+    });
+    if (!filePath) return;
+
+    try {
+      const project = await invoke<{
+        version: number;
+        name: string;
+        audio_file: string;
+        segments: Array<{
+          text: string;
+          start_ms: number;
+          end_ms: number;
+          speaker: string | null;
+          is_edited: boolean;
+          words: Array<{ word: string; start_ms: number; end_ms: number; confidence: number }>;
+        }>;
+        speakers: Array<{ label: string; display_name: string | null; color: string }>;
+      }>('load_project_file', { path: filePath });
+
+      // Rebuild speakers
+      const newSpeakers: Speaker[] = project.speakers.map((s, idx) => ({
+        id: `speaker-${idx}`,
+        project_id: '',
+        label: s.label,
+        display_name: s.display_name,
+        color: s.color,
+      }));
+      speakers.set(newSpeakers);
+
+      const speakerLookup = new Map(newSpeakers.map(s => [s.label, s.id]));
+
+      // Rebuild segments
+      const newSegments: Segment[] = project.segments.map((seg, idx) => ({
+        id: `seg-${idx}`,
+        project_id: '',
+        media_file_id: '',
+        speaker_id: seg.speaker ? (speakerLookup.get(seg.speaker) ?? null) : null,
+        start_ms: seg.start_ms,
+        end_ms: seg.end_ms,
+        text: seg.text,
+        original_text: null,
+        confidence: null,
+        is_edited: seg.is_edited,
+        edited_at: null,
+        segment_index: idx,
+        words: seg.words.map((w, widx) => ({
+          id: `word-${idx}-${widx}`,
+          segment_id: `seg-${idx}`,
+          word: w.word,
+          start_ms: w.start_ms,
+          end_ms: w.end_ms,
+          confidence: w.confidence,
+          word_index: widx,
+        })),
+      }));
+      segments.set(newSegments);
+
+      // Load audio
+      audioFilePath = project.audio_file;
+      audioUrl = convertFileSrc(project.audio_file);
+      waveformPlayer?.loadAudio(audioUrl);
+
+      currentProjectPath = filePath as string;
+      currentProjectName = project.name;
+    } catch (err) {
+      console.error('Failed to load project:', err);
+      alert(`Failed to load project: ${err}`);
+    }
+  }
+
   function handleWordClick(timeMs: number) {
+    console.log('[voice-to-notes] Word clicked, seeking to', timeMs, 'ms');
     waveformPlayer?.seekTo(timeMs);
   }
 
+  function handleTextEdit(segmentId: string, newText: string) {
+    // In-memory store is already updated by TranscriptEditor.
+    // Changes persist when user saves the project file.
+  }
+
   async function handleFileImport() {
     const filePath = await open({
       multiple: false,
@@ -81,14 +216,98 @@
     });
     if (!filePath) return;
 
-    // Convert file path to URL for wavesurfer
-    audioUrl = `asset://localhost/${encodeURIComponent(filePath)}`;
+    // Track the original file path and convert to asset URL for wavesurfer
+    audioFilePath = filePath;
+    audioUrl = convertFileSrc(filePath);
     waveformPlayer?.loadAudio(audioUrl);
 
+    // Clear previous results
+    segments.set([]);
+    speakers.set([]);
+
     // Start pipeline (transcription + diarization)
     isTranscribing = true;
     transcriptionProgress = 0;
     transcriptionStage = 'Starting...';
+    transcriptionMessage = 'Initializing pipeline...';
+
+    // Flush DOM so the progress overlay renders before the blocking invoke
+    await tick();
+
+    // Listen for progress events from the sidecar
+    const unlisten = await listen<{
+      percent: number;
+      stage: string;
+      message: string;
+    }>('pipeline-progress', (event) => {
+      console.log('[voice-to-notes] Progress event:', event.payload);
+      const { percent, stage, message } = event.payload;
+      if (typeof percent === 'number') transcriptionProgress = percent;
+      if (typeof stage === 'string') transcriptionStage = stage;
+      if (typeof message === 'string') transcriptionMessage = message;
+    });
+
+    const unlistenSegment = await listen<{
+      index: number;
+      text: string;
+      start_ms: number;
+      end_ms: number;
+      words: Array<{ word: string; start_ms: number; end_ms: number; confidence: number }>;
+    }>('pipeline-segment', (event) => {
+      const seg = event.payload;
+      const newSeg: Segment = {
+        id: `seg-${seg.index}`,
+        project_id: '',
+        media_file_id: '',
+        speaker_id: null,
+        start_ms: seg.start_ms,
+        end_ms: seg.end_ms,
+        text: seg.text,
+        original_text: null,
+        confidence: null,
+        is_edited: false,
+        edited_at: null,
+        segment_index: seg.index,
+        words: seg.words.map((w, widx) => ({
+          id: `word-${seg.index}-${widx}`,
+          segment_id: `seg-${seg.index}`,
+          word: w.word,
+          start_ms: w.start_ms,
+          end_ms: w.end_ms,
+          confidence: w.confidence,
+          word_index: widx,
+        })),
+      };
+      segments.update(segs => [...segs, newSeg]);
+    });
+
+    const unlistenSpeaker = await listen<{
+      updates: Array<{ index: number; speaker: string }>;
+    }>('pipeline-speaker-update', (event) => {
+      const { updates } = event.payload;
+      // Build speakers from unique labels
+      const uniqueLabels = [...new Set(updates.map(u => u.speaker))].sort();
+      const newSpeakers: Speaker[] = uniqueLabels.map((label, idx) => ({
+        id: `speaker-${idx}`,
+        project_id: '',
+        label,
+        display_name: null,
+        color: speakerColors[idx % speakerColors.length],
+      }));
+      speakers.set(newSpeakers);
+
+      // Update existing segments with speaker assignments
+      const speakerLookup = new Map(newSpeakers.map(s => [s.label, s.id]));
+      segments.update(segs =>
+        segs.map((seg, i) => {
+          const update = updates.find(u => u.index === i);
+          if (update) {
+            return { ...seg, speaker_id: speakerLookup.get(update.speaker) ?? null };
+          }
+          return seg;
+        })
+      );
+    });
 
     try {
       const result = await invoke<{
@@ -114,6 +333,8 @@
         device: $settings.transcription_device || undefined,
         language: $settings.transcription_language || undefined,
         skipDiarization: $settings.skip_diarization || undefined,
+        hfToken: $settings.hf_token || undefined,
+        numSpeakers: $settings.num_speakers && $settings.num_speakers > 0 ? $settings.num_speakers : undefined,
       });
 
       // Create speaker entries from pipeline result
@@ -155,10 +376,18 @@
       }));
 
       segments.set(newSegments);
+
+      // Set project name from audio file name (user can save explicitly)
+      const fileName = filePath.split(/[\\/]/).pop() || 'Untitled';
+      currentProjectName = fileName.replace(/\.[^.]+$/, '');
+      currentProjectPath = null;
     } catch (err) {
       console.error('Pipeline failed:', err);
       alert(`Pipeline failed: ${err}`);
     } finally {
+      unlisten();
+      unlistenSegment();
+      unlistenSpeaker();
       isTranscribing = false;
     }
   }
@@ -214,56 +443,77 @@
   }
 </script>
 
-<div class="app-header">
-  <h1>Voice to Notes</h1>
-  <div class="header-actions">
-    <button class="import-btn" onclick={handleFileImport}>
-      Import Audio/Video
-    </button>
-    <button class="settings-btn" onclick={() => showSettings = true} title="Settings">
-      Settings
-    </button>
-    {#if $segments.length > 0}
-      <div class="export-dropdown">
-        <button class="export-btn" onclick={() => showExportMenu = !showExportMenu}>
-          Export
+{#if !appReady}
+  <div class="splash-screen">
+    <h1 class="splash-title">Voice to Notes</h1>
+    <p class="splash-subtitle">Loading...</p>
+    <div class="splash-spinner"></div>
+  </div>
+{:else}
+  <div class="app-shell">
+  <div class="app-header">
+    <div class="header-actions">
+      <button class="settings-btn" onclick={openProject} disabled={isTranscribing}>
+        Open Project
+      </button>
+      {#if $segments.length > 0}
+        <button class="settings-btn" onclick={saveProject}>
+          Save Project
         </button>
-        {#if showExportMenu}
-          <div class="export-menu">
-            {#each exportFormats as fmt}
-              <button class="export-option" onclick={() => handleExport(fmt.format, fmt.ext, fmt.name)}>
-                {fmt.name} (.{fmt.ext})
-              </button>
-            {/each}
-          </div>
+      {/if}
+      <button class="import-btn" onclick={handleFileImport} disabled={isTranscribing}>
+        {#if isTranscribing}
+          Processing...
+        {:else}
+          Import Audio/Video
         {/if}
-      </div>
-    {/if}
+      </button>
+      <button class="settings-btn" onclick={() => showSettings = true} title="Settings">
+        Settings
+      </button>
+      {#if $segments.length > 0}
+        <div class="export-dropdown">
+          <button class="export-btn" onclick={() => showExportMenu = !showExportMenu}>
+            Export
+          </button>
+          {#if showExportMenu}
+            <div class="export-menu">
+              {#each exportFormats as fmt}
+                <button class="export-option" onclick={() => handleExport(fmt.format, fmt.ext, fmt.name)}>
+                  {fmt.name} (.{fmt.ext})
+                </button>
+              {/each}
+            </div>
+          {/if}
+        </div>
+      {/if}
+    </div>
   </div>
-</div>
 
-<div class="workspace">
-  <div class="main-content">
-    <WaveformPlayer bind:this={waveformPlayer} {audioUrl} />
-    <TranscriptEditor onWordClick={handleWordClick} />
+  <div class="workspace">
+    <div class="main-content">
+      <WaveformPlayer bind:this={waveformPlayer} {audioUrl} />
+      <TranscriptEditor onWordClick={handleWordClick} onTextEdit={handleTextEdit} />
+    </div>
+    <div class="sidebar-right">
+      <SpeakerManager />
+      <AIChatPanel />
+    </div>
   </div>
-  <div class="sidebar-right">
-    <SpeakerManager />
-    <AIChatPanel />
   </div>
-</div>
 
-<ProgressOverlay
-  visible={isTranscribing}
-  percent={transcriptionProgress}
-  stage={transcriptionStage}
-  message={transcriptionMessage}
-/>
+  <ProgressOverlay
+    visible={isTranscribing}
+    percent={transcriptionProgress}
+    stage={transcriptionStage}
+    message={transcriptionMessage}
+  />
 
-<SettingsModal
-  visible={showSettings}
-  onClose={() => showSettings = false}
-/>
+  <SettingsModal
+    visible={showSettings}
+    onClose={() => showSettings = false}
+  />
+{/if}
 
 <style>
   .app-header {
@@ -274,10 +524,6 @@
     background: #0f3460;
     color: #e0e0e0;
   }
-  h1 {
-    font-size: 1.25rem;
-    margin: 0;
-  }
   .import-btn {
     background: #e94560;
     border: none;
@@ -288,9 +534,18 @@
     font-size: 0.875rem;
     font-weight: 500;
   }
-  .import-btn:hover {
+  .import-btn:hover:not(:disabled) {
     background: #d63851;
   }
+  .import-btn:disabled {
+    opacity: 0.7;
+    cursor: not-allowed;
+    animation: pulse 1.5s ease-in-out infinite;
+  }
+  @keyframes pulse {
+    0%, 100% { opacity: 0.7; }
+    50% { opacity: 1; }
+  }
   .header-actions {
     display: flex;
     gap: 0.5rem;
@@ -305,10 +560,14 @@
     cursor: pointer;
     font-size: 0.875rem;
   }
-  .settings-btn:hover {
+  .settings-btn:hover:not(:disabled) {
     background: rgba(255,255,255,0.05);
     border-color: #e94560;
   }
+  .settings-btn:disabled {
+    opacity: 0.5;
+    cursor: not-allowed;
+  }
   .export-dropdown {
     position: relative;
   }
@@ -351,11 +610,19 @@
   .export-option:hover {
     background: rgba(233, 69, 96, 0.2);
   }
+  .app-shell {
+    display: flex;
+    flex-direction: column;
+    height: 100vh;
+    overflow: hidden;
+  }
   .workspace {
     display: flex;
     gap: 1rem;
     padding: 1rem;
-    height: calc(100vh - 3.5rem);
+    flex: 1;
+    min-height: 0;
+    overflow: hidden;
     background: #0a0a23;
   }
   .main-content {
@@ -364,6 +631,8 @@
     flex-direction: column;
     gap: 1rem;
     min-width: 0;
+    min-height: 0;
+    overflow-y: auto;
   }
   .sidebar-right {
     width: 300px;
@@ -371,5 +640,38 @@
     flex-direction: column;
     gap: 1rem;
     flex-shrink: 0;
+    min-height: 0;
+    overflow-y: auto;
+  }
+  .splash-screen {
+    display: flex;
+    flex-direction: column;
+    align-items: center;
+    justify-content: center;
+    height: 100vh;
+    background: #0a0a23;
+    color: #e0e0e0;
+    gap: 1rem;
+  }
+  .splash-title {
+    font-size: 2rem;
+    margin: 0;
+    color: #e94560;
+  }
+  .splash-subtitle {
+    font-size: 1rem;
+    color: #888;
+    margin: 0;
+  }
+  .splash-spinner {
+    width: 32px;
+    height: 32px;
+    border: 3px solid #2a3a5e;
+    border-top-color: #e94560;
+    border-radius: 50%;
+    animation: spin 0.8s linear infinite;
+  }
+  @keyframes spin {
+    to { transform: rotate(360deg); }
   }
 </style>