deepinfra.py

transcription_pipeline/engines/deepinfra.py

1. """DeepInfra's hosted Whisper large-v3-turbo.
3. The audio leaves the machine for this step only; alignment and speaker
4. separation still run locally.
5. """
7. from __future__ import annotations
9. import json
10. import urllib.error
11. import urllib.request
12. import uuid
13. from pathlib import Path
15. API_URL = "https://api.deepinfra.com/v1/inference/openai/whisper-large-v3-turbo"
16. MODEL_LABEL = "openai/whisper-large-v3-turbo (DeepInfra)"
19. class EngineError(RuntimeError):
20. pass
23. def _multipart(fields: dict[str, str], file_field: str, path: Path) -> tuple[bytes, str]:
24. boundary = uuid.uuid4().hex
25. parts = []
26. for k, v in fields.items():
27. parts.append((f'--{boundary}\r\nContent-Disposition: form-data; name="{k}"\r\n\r\n'
28. f'{v}\r\n').encode())
29. parts.append((f'--{boundary}\r\nContent-Disposition: form-data; name="{file_field}"; '
30. f'filename="{path.name}"\r\n'
31. 'Content-Type: application/octet-stream\r\n\r\n').encode())
32. parts.append(path.read_bytes())
33. parts.append(f"\r\n--{boundary}--\r\n".encode())
34. return b"".join(parts), f"multipart/form-data; boundary={boundary}"
37. def transcribe(audio_path: Path, token: str | None, language: str | None = None,
38. api_url: str = API_URL, timeout: float = 900) -> dict:
39. if not token:
40. raise EngineError("No DeepInfra token: set DEEPINFRA_TOKEN or add it to the Keychain "
41. "(see docs/SETUP.md).")
42. fields = {"language": language} if language else {}
43. body, ctype = _multipart(fields, "audio", audio_path)
44. headers = {"Authorization": f"bearer {token}", "Content-Type": ctype}
45. req = urllib.request.Request(api_url, data=body, method="POST", headers=headers)
46. try:
47. with urllib.request.urlopen(req, timeout=timeout) as res:
48. data = json.loads(res.read().decode("utf-8"))
49. except urllib.error.HTTPError as e:
50. if e.code in (401, 403):
51. raise EngineError(
52. "DeepInfra rejected the token: it has been revoked or rotated.") from e
53. if e.code == 429:
54. raise EngineError("DeepInfra is rate limiting, or the account is out of credit.") from e
55. if e.code == 413:
56. raise EngineError("The file is too large for DeepInfra.") from e
57. detail = e.read().decode("utf-8", "replace")[:300]
58. raise EngineError(f"DeepInfra error {e.code}: {detail}") from e
59. except urllib.error.URLError as e:
60. raise EngineError(f"Could not reach DeepInfra: {e.reason}") from e
62. segments = [{"start": float(s.get("start") or 0), "end": float(s.get("end") or 0),
63. "text": (s.get("text") or "").strip()}
64. for s in (data.get("segments") or []) if (s.get("text") or "").strip()]
65. if not segments and (data.get("text") or "").strip():
66. segments = [{"start": 0.0, "end": 0.0, "text": data["text"].strip()}]
67. return {"segments": segments, "language": data.get("language") or language,
68. "model": MODEL_LABEL}