103 lines
3.9 KiB
Python
103 lines
3.9 KiB
Python
import mlx_whisper
|
|
import numpy as np
|
|
import sounddevice as sd
|
|
import queue
|
|
import sys
|
|
import torch
|
|
from silero_vad import load_silero_vad, get_speech_timestamps
|
|
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
|
|
|
|
# Parameters
|
|
WHISPER_MODEL = "mlx-community/whisper-small.en-mlx"
|
|
TRANSLATE_MODEL = "facebook/nllb-200-distilled-600M"
|
|
TARGET_LANG = "spa_Latn" # Spanish (Latin America) - change as needed
|
|
CHANNELS = 1
|
|
SAMPLERATE = 16000
|
|
BLOCK_SIZE = 512
|
|
VAD_THRESHOLD = 0.5
|
|
BUFFER_LIMIT = SAMPLERATE * 30
|
|
MIN_SILENCE_DURATION_MS = 500
|
|
|
|
audio_queue = queue.Queue()
|
|
|
|
def callback(indata, frames, time, status):
|
|
if status:
|
|
print(status, file=sys.stderr)
|
|
audio_queue.put(indata.copy())
|
|
|
|
def main():
|
|
# Set device for translation model (using MPS for Mac M chips)
|
|
device = "mps" if torch.backends.mps.is_available() else "cpu"
|
|
print(f"Using device: {device}")
|
|
|
|
print(f"Loading Whisper model '{WHISPER_MODEL}'...")
|
|
# Whisper MLX runs on its own optimized path
|
|
|
|
print(f"Loading translation model '{TRANSLATE_MODEL}'...")
|
|
tokenizer = AutoTokenizer.from_pretrained(TRANSLATE_MODEL)
|
|
model = AutoModelForSeq2SeqLM.from_pretrained(TRANSLATE_MODEL).to(device)
|
|
translator = pipeline("translation", model=model, tokenizer=tokenizer,
|
|
src_lang="eng_Latn", tgt_lang=TARGET_LANG, device=device)
|
|
|
|
print("Loading Silero VAD model...")
|
|
vad_model = load_silero_vad()
|
|
|
|
print("Models loaded.")
|
|
|
|
print(f"\nStarting live transcription & translation to {TARGET_LANG}... (Press Ctrl+C to stop)")
|
|
|
|
audio_buffer = []
|
|
speech_started = False
|
|
|
|
try:
|
|
with sd.InputStream(samplerate=SAMPLERATE, channels=CHANNELS, callback=callback, blocksize=BLOCK_SIZE):
|
|
while True:
|
|
while not audio_queue.empty():
|
|
data = audio_queue.get()
|
|
audio_buffer.append(data.flatten())
|
|
|
|
if len(audio_buffer) > 0:
|
|
current_audio = np.concatenate(audio_buffer)
|
|
audio_tensor = torch.from_numpy(current_audio)
|
|
|
|
speech_timestamps = get_speech_timestamps(
|
|
audio_tensor,
|
|
vad_model,
|
|
sampling_rate=SAMPLERATE,
|
|
threshold=VAD_THRESHOLD,
|
|
min_silence_duration_ms=MIN_SILENCE_DURATION_MS
|
|
)
|
|
|
|
if len(speech_timestamps) > 0:
|
|
speech_started = True
|
|
last_end = speech_timestamps[-1]['end']
|
|
buffer_len_samples = len(current_audio)
|
|
|
|
if (buffer_len_samples - last_end) > (SAMPLERATE * MIN_SILENCE_DURATION_MS / 1000) or buffer_len_samples > BUFFER_LIMIT:
|
|
|
|
# 1. Transcribe (English)
|
|
result = mlx_whisper.transcribe(current_audio, path_or_hf_repo=WHISPER_MODEL)
|
|
original_text = result['text'].strip()
|
|
|
|
if original_text:
|
|
# 2. Translate
|
|
translation = translator(original_text)
|
|
translated_text = translation[0]['translation_text']
|
|
|
|
print(f"\nEN: {original_text}")
|
|
print(f"ES: {translated_text}")
|
|
|
|
audio_buffer = []
|
|
speech_started = False
|
|
|
|
elif not speech_started and len(current_audio) > SAMPLERATE * 2:
|
|
audio_buffer = []
|
|
|
|
except KeyboardInterrupt:
|
|
print("\nStopped by user.")
|
|
except Exception as e:
|
|
print(f"\nError: {e}")
|
|
|
|
if __name__ == "__main__":
|
|
main()
|