Optimize local whisper

src/server/backends/whisper.js

@@ -16,6 +16,39 @@ function getAudioThreshold() {
   return settingsManager.getSetting('audioThreshold') || parseFloat(process.env.AUDIO_THRESHOLD) || 0.01;
 }
 
+// Preload tiny model for faster initial transcriptions
+async function warmupWhisper() {
+  console.log('Warming up Whisper with tiny model...');
+  const tempDir = os.tmpdir();
+  const tempFile = path.join(tempDir, 'warmup_audio.wav');
+  
+  try {
+    // Create a minimal audio file for warmup
+    const silentAudio = Buffer.alloc(8000); // 0.5 seconds of silence
+    await fs.writeFile(tempFile, silentAudio);
+    
+    const warmupProcess = spawn('python', [
+      '-m', 'whisper', tempFile, 
+      '--model', 'tiny', 
+      '--language', 'en',
+      '--output_format', 'txt',
+      '--output_dir', tempDir
+    ]);
+    
+    warmupProcess.on('close', async () => {
+      await fs.unlink(tempFile).catch(() => {});
+      await fs.unlink(tempFile.replace('.wav', '.txt')).catch(() => {});
+      console.log('Whisper warmup completed');
+    });
+    
+  } catch (error) {
+    console.log('Whisper warmup failed (optional):', error.message);
+  }
+}
+
+// Warmup on module load
+setTimeout(warmupWhisper, 2000); // Delay to let server start first
+
 async function transcribeLocal(audioBuffer) {
   // Check audio level before processing
   const audioLevel = getAudioLevel(audioBuffer);
@@ -37,15 +70,23 @@ async function transcribeLocal(audioBuffer) {
       await fs.writeFile(tempFile, audioBuffer);
       
       return new Promise((resolve, reject) => {
-        // Build Whisper command with current settings (not cached)
-        const currentModel = settingsManager.getSetting('whisperModel') || process.env.WHISPER_MODEL || 'base';
+        // Build Whisper command with current settings (optimized for speed)
+        const selectedModel = settingsManager.getSetting('whisperModel') || process.env.WHISPER_MODEL || 'base';
+        
+        // Use tiny model for faster processing of short audio chunks
+        const currentModel = audioBuffer.length < 32000 ? 'tiny' : selectedModel; // ~2 seconds of audio
+        
         const whisperArgs = [
           tempFile,
           '--model', currentModel,
           '--language', 'en',
           '--task', 'transcribe',
           '--output_format', 'txt',
-          '--output_dir', tempDir
+          '--output_dir', tempDir,
+          '--no_speech_threshold', '0.6',  // Skip processing obvious non-speech
+          '--logprob_threshold', '-1.0',   // Skip low-confidence segments
+          '--compression_ratio_threshold', '2.4',  // Skip repetitive audio
+          '--condition_on_previous_text', 'False'  // Don't wait for previous context
         ];
 
         // Get current GPU settings (not cached)