|
1
|
"""Local Whisper transcription provider — runs on-device with GPU acceleration.""" |
|
2
|
|
|
3
|
import logging |
|
4
|
from pathlib import Path |
|
5
|
from typing import Optional |
|
6
|
|
|
7
|
logger = logging.getLogger(__name__) |
|
8
|
|
|
9
|
# Model size → approximate VRAM/RAM usage |
|
10
|
_MODEL_SIZES = { |
|
11
|
"tiny": "~1GB", |
|
12
|
"base": "~1GB", |
|
13
|
"small": "~2GB", |
|
14
|
"medium": "~5GB", |
|
15
|
"large": "~10GB", |
|
16
|
"turbo": "~6GB", |
|
17
|
} |
|
18
|
|
|
19
|
|
|
20
|
class WhisperLocal: |
|
21
|
""" |
|
22
|
Local Whisper transcription using openai-whisper. |
|
23
|
|
|
24
|
Uses MPS (Apple Silicon) or CUDA when available, falls back to CPU. |
|
25
|
No file size limits — processes audio directly on device. |
|
26
|
""" |
|
27
|
|
|
28
|
def __init__(self, model_size: str = "large", device: Optional[str] = None): |
|
29
|
""" |
|
30
|
Initialize local Whisper. |
|
31
|
|
|
32
|
Parameters |
|
33
|
---------- |
|
34
|
model_size : str |
|
35
|
Whisper model size: tiny, base, small, medium, large, turbo |
|
36
|
device : str, optional |
|
37
|
Force device: 'mps', 'cuda', 'cpu'. Auto-detects if None. |
|
38
|
""" |
|
39
|
self.model_size = model_size |
|
40
|
self._model = None |
|
41
|
|
|
42
|
if device: |
|
43
|
self.device = device |
|
44
|
else: |
|
45
|
self.device = self._detect_device() |
|
46
|
|
|
47
|
logger.info( |
|
48
|
f"WhisperLocal: model={model_size} ({_MODEL_SIZES.get(model_size, '?')}), " |
|
49
|
f"device={self.device}" |
|
50
|
) |
|
51
|
|
|
52
|
@staticmethod |
|
53
|
def _detect_device() -> str: |
|
54
|
"""Auto-detect the best available device.""" |
|
55
|
try: |
|
56
|
import torch |
|
57
|
|
|
58
|
if torch.cuda.is_available(): |
|
59
|
return "cuda" |
|
60
|
if torch.backends.mps.is_available(): |
|
61
|
return "mps" |
|
62
|
except ImportError: |
|
63
|
pass |
|
64
|
return "cpu" |
|
65
|
|
|
66
|
def _load_model(self): |
|
67
|
"""Lazy-load the Whisper model.""" |
|
68
|
if self._model is not None: |
|
69
|
return |
|
70
|
|
|
71
|
try: |
|
72
|
import whisper |
|
73
|
except ImportError: |
|
74
|
raise ImportError("openai-whisper not installed. Run: pip install openai-whisper torch") |
|
75
|
|
|
76
|
logger.info(f"Loading Whisper {self.model_size} model on {self.device}...") |
|
77
|
self._model = whisper.load_model(self.model_size, device=self.device) |
|
78
|
logger.info("Whisper model loaded") |
|
79
|
|
|
80
|
def transcribe( |
|
81
|
self, |
|
82
|
audio_path: str | Path, |
|
83
|
language: Optional[str] = None, |
|
84
|
) -> dict: |
|
85
|
""" |
|
86
|
Transcribe audio using local Whisper. |
|
87
|
|
|
88
|
No file size limits. Runs entirely on device. |
|
89
|
|
|
90
|
Returns dict compatible with ProviderManager transcription format. |
|
91
|
""" |
|
92
|
self._load_model() |
|
93
|
audio_path = Path(audio_path) |
|
94
|
|
|
95
|
logger.info(f"Transcribing {audio_path.name} with Whisper {self.model_size}...") |
|
96
|
|
|
97
|
# fp16 only works reliably on CUDA; MPS produces NaN with large models |
|
98
|
kwargs = {"fp16": self.device == "cuda"} |
|
99
|
if language: |
|
100
|
kwargs["language"] = language |
|
101
|
|
|
102
|
result = self._model.transcribe(str(audio_path), **kwargs) |
|
103
|
|
|
104
|
segments = [ |
|
105
|
{ |
|
106
|
"start": seg["start"], |
|
107
|
"end": seg["end"], |
|
108
|
"text": seg["text"].strip(), |
|
109
|
} |
|
110
|
for seg in result.get("segments", []) |
|
111
|
] |
|
112
|
|
|
113
|
duration = segments[-1]["end"] if segments else None |
|
114
|
|
|
115
|
return { |
|
116
|
"text": result.get("text", "").strip(), |
|
117
|
"segments": segments, |
|
118
|
"language": result.get("language", language), |
|
119
|
"duration": duration, |
|
120
|
"provider": "whisper-local", |
|
121
|
"model": f"whisper-{self.model_size}", |
|
122
|
} |
|
123
|
|
|
124
|
@staticmethod |
|
125
|
def is_available() -> bool: |
|
126
|
"""Check if local Whisper is installed and usable.""" |
|
127
|
try: |
|
128
|
import torch # noqa: F401 |
|
129
|
import whisper # noqa: F401 |
|
130
|
|
|
131
|
return True |
|
132
|
except ImportError: |
|
133
|
return False |
|
134
|
|