torchaudio 2.10 unconditionally delegates load() to torchcodec, ignoring
the backend parameter. Since torchcodec is excluded from PyInstaller,
this broke our pyannote Audio monkey-patch.
Fix: replace torchaudio.load() with soundfile.read() + torch.from_numpy().
soundfile handles WAV natively (audio is pre-converted to WAV), has no
torchcodec dependency, and is already a transitive dependency.
Also added soundfile to PyInstaller hiddenimports.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
torchcodec is partially bundled but non-functional (missing FFmpeg DLLs),
causing pyannote.audio to try AudioDecoder which fails with NameError.
Excluding it forces pyannote to fall back to torchaudio for audio loading.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Exclude ctranslate2.converters from PyInstaller bundle — these modules
import torch at module level causing circular import crashes, and are
only needed for model conversion (never used at runtime)
- Defer all heavy ML imports to first handler call instead of startup,
so the sidecar can send its ready message without loading torch/whisper
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
