adolforeyna/whisper-translation
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix variable initialization order (vad_model defined before stream start)

Browse Source
This commit is contained in:
Adolfo Reyna
2026-02-26 21:34:07 -05:00
parent d75e19665f
commit da73d4e0c7
1 changed files with 10 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
transcribe.py
View File

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