Fix variable initialization order (vad_model defined before stream start)

transcribe.py

@@ -51,20 +51,23 @@ def main():
     device = "mps" if torch.backends.mps.is_available() else "cpu"
     print(f"Using device: {device}")
 
+    # 1. Load models first
     print(f"Loading Whisper model '{WHISPER_MODEL}'...")
     
-    # Dictionary to hold models and tokenizers
     translation_engines = {}
-    
     for lang_name, model_id in TARGET_LANGS.items():
         print(f"Loading {lang_name} translation model ({model_id})...")
         tokenizer = MarianTokenizer.from_pretrained(model_id)
         model = MarianMTModel.from_pretrained(model_id).to(device)
         translation_engines[lang_name] = (model, tokenizer)
 
+    print("Loading Silero VAD model...")
+    vad_model = load_silero_vad()
+    print("Models loaded.")
+
+    # 2. Select Audio Device
     print("\nAvailable Audio Devices:")
-    devices = sd.query_devices()
-    print(devices)
+    print(sd.query_devices())
     
     try:
         device_input = input("\nSelect input device index (or press Enter for default): ")
@@ -79,6 +82,7 @@ def main():
     speech_started = False
 
     try:
+        # 3. Start stream
         with sd.InputStream(samplerate=SAMPLERATE, channels=CHANNELS, callback=callback, blocksize=BLOCK_SIZE, device=device_index):
             while True:
                 while not audio_queue.empty():
@@ -104,22 +108,20 @@ def main():
                         
                         if (buffer_len_samples - last_end) > (SAMPLERATE * MIN_SILENCE_DURATION_MS / 1000) or buffer_len_samples > BUFFER_LIMIT:
                             
-                            # 1. Transcribe once
+                            # Transcribe once
                             result = mlx_whisper.transcribe(current_audio, path_or_hf_repo=WHISPER_MODEL)
                             original_text = result['text'].strip()
                             
                             if original_text:
-                                # Limit the input length to avoid memory spikes or model glitches
                                 if len(original_text) > 250:
                                     original_text = original_text[:247] + "..."
                                 
                                 print(f"\n[EN]: {original_text}")
                                 
-                                # 2. Translate to all targets
+                                # Translate to all targets
                                 for lang_name, (model, tokenizer) in translation_engines.items():
                                     inputs = tokenizer(original_text, return_tensors="pt", padding=True).to(device)
                                     with torch.no_grad():
-                                        # Added max_new_tokens to prevent runaway generation
                                         translated_tokens = model.generate(**inputs, max_new_tokens=150)
                                     translated_text = tokenizer.decode(translated_tokens[0], skip_special_tokens=True)
                                     print(f"[{lang_name[:2].upper()}]: {translated_text}")