Improve transcription with Silero VAD for better speech detection

transcribe.py

@@ -3,13 +3,17 @@ import numpy as np
 import sounddevice as sd
 import queue
 import sys
+import torch
+from silero_vad import load_silero_vad, get_speech_timestamps
 
 # Parameters
 MODEL_TYPE = "tiny.en"
 CHANNELS = 1
 SAMPLERATE = 16000
-BLOCK_SIZE = 8000 # 0.5 seconds of audio per block
-TRANSCRIBE_RATE = 2 # Process every 2 seconds
+BLOCK_SIZE = 512 # Silero VAD prefers specific block sizes (512, 1024, 1536)
+VAD_THRESHOLD = 0.5 # Confidence threshold for speech
+BUFFER_LIMIT = SAMPLERATE * 30 # Max 30 seconds of audio buffer
+MIN_SILENCE_DURATION_MS = 500 # Silence duration to trigger transcription
 
 audio_queue = queue.Queue()
 
@@ -20,42 +24,74 @@ def callback(indata, frames, time, status):
 
 def main():
     print(f"Loading Whisper model '{MODEL_TYPE}'...")
-    model = whisper.load_model(MODEL_TYPE)
-    print("Model loaded.")
+    whisper_model = whisper.load_model(MODEL_TYPE)
+    
+    print("Loading Silero VAD model...")
+    vad_model = load_silero_vad()
+    
+    print("Models loaded.")
 
     print("\nAvailable Audio Devices:")
     devices = sd.query_devices()
     print(devices)
     
-    # Try to find a sensible default if the system one is tricky
     default_device = sd.default.device[0]
     print(f"\nUsing default input device index: {default_device}")
 
-    print("\nStarting live transcription... (Press Ctrl+C to stop)")
-    print("Note: On macOS, you may need to grant Microphone permissions to your terminal.\n")
+    print("\nStarting live transcription with VAD... (Press Ctrl+C to stop)")
     
-    audio_buffer = np.array([], dtype=np.float32)
+    audio_buffer = []
+    speech_started = False
 
     try:
         with sd.InputStream(samplerate=SAMPLERATE, channels=CHANNELS, callback=callback, blocksize=BLOCK_SIZE):
             while True:
-                # Pull all available data from the queue
                 while not audio_queue.empty():
                     data = audio_queue.get()
-                    audio_buffer = np.append(audio_buffer, data.flatten())
+                    audio_buffer.append(data.flatten())
 
-                # If we have enough audio, transcribe it
-                if len(audio_buffer) >= SAMPLERATE * TRANSCRIBE_RATE:
-                    # Transcribe the current buffer
-                    # fp16=False is used for CPU execution
-                    result = model.transcribe(audio_buffer, fp16=False, language="en")
+                if len(audio_buffer) > 0:
+                    # Concatenate buffer to check for speech
+                    current_audio = np.concatenate(audio_buffer)
+                    
+                    # Convert to torch tensor for Silero
+                    audio_tensor = torch.from_numpy(current_audio)
+                    
+                    # Get speech timestamps
+                    speech_timestamps = get_speech_timestamps(
+                        audio_tensor, 
+                        vad_model, 
+                        sampling_rate=SAMPLERATE,
+                        threshold=VAD_THRESHOLD,
+                        min_silence_duration_ms=MIN_SILENCE_DURATION_MS
+                    )
+
+                    # If we have speech and then silence, or buffer is getting too long
+                    if len(speech_timestamps) > 0:
+                        speech_started = True
+                        
+                        # Check if the last speech segment has "ended" (i.e., we have enough silence after it)
+                        # or if we've reached a significant buffer size
+                        last_end = speech_timestamps[-1]['end']
+                        buffer_len_samples = len(current_audio)
+                        
+                        # If the speech ended more than MIN_SILENCE_DURATION_MS ago
+                        if (buffer_len_samples - last_end) > (SAMPLERATE * MIN_SILENCE_DURATION_MS / 1000) or buffer_len_samples > BUFFER_LIMIT:
+                            
+                            # Transcribe the valid speech segment
+                            result = whisper_model.transcribe(current_audio, fp16=False, language="en")
                             text = result['text'].strip()
                             
                             if text:
                                 print(f"Transcription: {text}")
                             
-                    # Clear buffer for next chunk
-                    audio_buffer = np.array([], dtype=np.float32)
+                            # Reset buffer
+                            audio_buffer = []
+                            speech_started = False
+                    
+                    elif not speech_started and len(current_audio) > SAMPLERATE * 2:
+                        # Clear buffer if it's just silence for more than 2 seconds
+                        audio_buffer = []
 
     except KeyboardInterrupt:
         print("\nStopped by user.")