steps.py

transcription_pipeline/steps.py

1. """The real model-backed steps, wired to pipeline.process's callables.
3. Imports of WhisperX and torch happen inside the steps, so the rest of the
4. package (and its tests) run without them.
5. """
7. from __future__ import annotations
9. import tempfile
10. from pathlib import Path
12. from . import audio
13. from .secrets import get_secret
16. def transcriber(opts):
17. if opts.engine == "deepinfra":
18. from .engines import deepinfra
19. token = get_secret("DEEPINFRA_TOKEN")
21. def run(wav: Path, source: Path, o):
22. # Upload FLAC, not the WAV: same audio, about half the bytes.
23. with tempfile.TemporaryDirectory(prefix="tp-up-") as tmp:
24. flac = audio.to_flac16k(wav, Path(tmp) / "audio.flac")
25. return deepinfra.transcribe(flac, token, language=o.language)
26. return run
27. if opts.engine == "local":
28. from .engines import local
29. return lambda wav, source, o: local.transcribe(wav, language=o.language, model=o.model)
30. raise ValueError(f"Unknown engine {opts.engine!r}: use local or deepinfra.")
33. def aligner(opts):
34. from .align import align
35. return lambda result, wav, o: align(result, wav)
38. def diarizer(opts):
39. if not opts.diarize:
40. return None
41. from .diarize import diarize
42. token = get_secret("HF_TOKEN")
43. return lambda result, wav, o: diarize(result, wav, token, o.min_speakers, o.max_speakers)