local.py

transcription_pipeline/engines/local.py

1. """Local transcription: faster-whisper large-v3-turbo through WhisperX.
3. The default engine. Nothing leaves the machine. int8 on the CPU is the setting
4. that runs everywhere, Apple Silicon included (CTranslate2 has no Metal backend,
5. so there is no GPU path on a Mac for this step).
6. """
8. from __future__ import annotations
10. from functools import lru_cache
11. from pathlib import Path
13. from ..audio import load_wav16k
15. DEFAULT_MODEL = "large-v3-turbo"
18. @lru_cache(maxsize=2)
19. def _model(name: str, device: str, compute_type: str, language: str | None):
20. import whisperx
21. return whisperx.load_model(name, device, compute_type=compute_type, language=language)
24. def transcribe(wav: Path, language: str | None = None, model: str = DEFAULT_MODEL,
25. device: str = "cpu", compute_type: str = "int8", batch_size: int = 8) -> dict:
26. audio = load_wav16k(wav)
27. result = _model(model, device, compute_type, language).transcribe(
28. audio, batch_size=batch_size, language=language)
29. segments = [{"start": float(s["start"]), "end": float(s["end"]), "text": s["text"].strip()}
30. for s in result.get("segments", []) if s.get("text", "").strip()]
31. return {"segments": segments, "language": result.get("language") or language,
32. "model": f"faster-whisper {model} ({compute_type}, {device})"}