Repository Public
1"""Transcribes one audio file with word-level timestamps (mlx-whisper, runs locally on Apple Silicon).23 python transcribe.py <audio file> <model repo> <language | auto>45Prints one JSON object on stdout: {"language": "fr", "segments": [{start, end, text, no_speech_prob, avg_logprob,6compression_ratio, words: [{start, end, word}]}]}. Called by the worker with an argument vector (no shell).7"""8import json9import math10import sys111213def num(value, fallback: float) -> float:14 """Whisper yields NaN for a segment with no tokens (silence, pure instrumental); NaN is not valid JSON."""15 try:16 f = float(value)17 except (TypeError, ValueError):18 return fallback19 return f if math.isfinite(f) else fallback202122def main() -> None:23 path, model, language = sys.argv[1], sys.argv[2], sys.argv[3]24 import mlx_whisper # imported late so a missing install gives a clear error line2526 result = mlx_whisper.transcribe(27 path,28 path_or_hf_repo=model,29 word_timestamps=True,30 # Songs repeat themselves: feeding the previous window back makes the model loop on a chorus.31 condition_on_previous_text=False,32 language=None if language == "auto" else language,33 verbose=None,34 )35 out = {"language": result.get("language"), "segments": []}36 for s in result.get("segments", []):37 out["segments"].append({38 "start": float(s["start"]), "end": float(s["end"]), "text": s["text"],39 "no_speech_prob": num(s.get("no_speech_prob"), 0.0), "avg_logprob": num(s.get("avg_logprob"), -5.0),40 "compression_ratio": num(s.get("compression_ratio"), 0.0),41 "words": [{"start": float(w["start"]), "end": float(w["end"]), "word": w["word"]} for w in s.get("words", [])],42 })43 json.dump(out, sys.stdout, ensure_ascii=False, allow_nan=False)444546if __name__ == "__main__":47 try:48 main()49 except Exception as e: # noqa: BLE001 — one line for the worker's log, no traceback with paths50 print(f"transcribe failed: {type(e).__name__}: {str(e)[:200]}", file=sys.stderr)51 sys.exit(1)52