Spaces:
Sleeping
Sleeping
File size: 2,233 Bytes
74bb5fe ac1f51b 74bb5fe ac1f51b 74bb5fe ac1f51b 74bb5fe |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 |
# models/asr_whisper.py
from __future__ import annotations
import os
from typing import Optional, Dict, Any
from faster_whisper import WhisperModel
from utils.config import get_settings
_asr_singleton: Optional["WhisperASR"] = None
def _norm_device(req: str) -> str:
"""faster-whisper supports only 'cpu' or 'cuda'."""
r = (req or "cpu").strip().lower()
if r == "mps":
print("[ASR] 'mps' is not supported by faster-whisper; falling back to CPU.")
return "cpu"
return r if r in ("cpu", "cuda") else "cpu"
def _compute_type_for(device: str) -> str:
# Keep it simple and stable for HF/macOS:
# - CPU: int8 (fast, small)
# - CUDA: float16 (good default on GPUs)
return "float16" if device == "cuda" else "int8"
class WhisperASR:
def __init__(self):
s = get_settings()
self.model_size = os.getenv("WHISPER_SIZE", "tiny").strip() # tiny|base|small|medium|large-v3 ...
self.language = os.getenv("WHISPER_LANG", "").strip() or None # e.g., "en"; None = auto
self.device = _norm_device(getattr(s, "ASR_DEVICE", "cpu"))
self.compute_type = _compute_type_for(self.device)
print(f"[ASR] Loading faster-whisper: size={self.model_size} device={self.device} compute_type={self.compute_type}")
self.model = WhisperModel(self.model_size, device=self.device, compute_type=self.compute_type)
def transcribe(self, path: str) -> Dict[str, Any]:
"""
Returns: {"text": str, "language": str|None, "segments": [...]}
"""
# language=None lets faster-whisper auto-detect. You can force via WHISPER_LANG.
segments, info = self.model.transcribe(
path,
beam_size=1,
language=self.language or None,
)
text = " ".join((seg.text or "").strip() for seg in segments).strip()
return {
"text": text or "",
"language": getattr(info, "language", None),
# You can expose timings later if you want:
"segments": [] # keep lightweight for UI
}
def get_asr() -> WhisperASR:
global _asr_singleton
if _asr_singleton is None:
_asr_singleton = WhisperASR()
return _asr_singleton |