TalkEdit/transcribe.py

#!/usr/bin/env python3
import sys
import json
import tempfile
import subprocess
from faster_whisper import WhisperModel

def extract_audio(input_path, output_path):
    """Extract audio from video/audio file to 16kHz mono WAV"""
    cmd = [
        'ffmpeg', '-y', '-i', input_path, '-vn', '-ar', '16000', '-ac', '1', '-f', 'wav', output_path
    ]
    subprocess.run(cmd, check=True)

def main():
    if len(sys.argv) < 3:
        print("Usage: python transcribe.py <audio_file> <model_name> [language]", file=sys.stderr)
        sys.exit(1)

    audio_file = sys.argv[1]
    model_name = sys.argv[2]
    language = sys.argv[3] if len(sys.argv) > 3 else None

    # Check file size - warn for very large files
    import os
    file_size_mb = os.path.getsize(audio_file) / (1024 * 1024)
    if file_size_mb > 100:  # Warn for files over 100MB
        print(f"Warning: Large file detected ({file_size_mb:.1f}MB). Transcription may take a long time.", file=sys.stderr)
        print("Consider splitting long audio files into smaller segments for faster processing.", file=sys.stderr)

    # Extract audio to temp WAV if needed
    with tempfile.NamedTemporaryFile(suffix='.wav', delete=False) as tmp:
        wav_path = tmp.name

    try:
        extract_audio(audio_file, wav_path)

        # Load model - use GPU if CUDA is available, else CPU with int8
        import ctypes
        try:
            ctypes.CDLL("libcublas.so.12")
            device = "cuda"
            compute_type = "float16"
        except OSError:
            device = "cpu"
            compute_type = "int8"

        model = WhisperModel(model_name, device=device, compute_type=compute_type)

        # Transcribe with progress reporting
        print(f"Starting transcription of {wav_path} with model {model_name}", file=sys.stderr)
        
        segments, info = model.transcribe(
            wav_path,
            language=language,
            beam_size=5,
            word_timestamps=True,
            vad_filter=True,
            vad_parameters=dict(threshold=0.5, min_speech_duration_ms=250),
            without_timestamps=False
        )
        
        print(f"Transcription completed. Detected language: {info.language}", file=sys.stderr)

        # Convert to our format
        words = []
        segments_list = []

        for segment in segments:
            seg_words = []
            for word in segment.words:
                w = {
                    "word": word.word,
                    "start": word.start,
                    "end": word.end,
                    "confidence": word.probability
                }
                words.append(w)
                seg_words.append(w)

            segments_list.append({
                "id": len(segments_list),
                "start": segment.start,
                "end": segment.end,
                "text": segment.text,
                "words": seg_words
            })

        result = {
            "words": words,
            "segments": segments_list,
            "language": info.language
        }

        print(json.dumps(result))

    finally:
        import os
        os.unlink(wav_path)

if __name__ == "__main__":
    main()