Spaces:
Running on Zero
Running on Zero
| """ | |
| UTMOS naturalness evaluator. | |
| Matches the F5-TTS evaluation approach: SpeechMOS UTMOS22 strong predictor | |
| loaded through torch.hub. Kept independent and lazy. | |
| """ | |
| import gc | |
| import threading | |
| from pathlib import Path | |
| import numpy as np | |
| import soundfile as sf | |
| import torch | |
| _LOCK = threading.RLock() | |
| _PREDICTOR = None | |
| _DEVICE = "cpu" | |
| def _load(): | |
| global _PREDICTOR | |
| with _LOCK: | |
| if _PREDICTOR is not None: | |
| return _PREDICTOR | |
| _PREDICTOR = torch.hub.load( | |
| "tarepan/SpeechMOS:v1.2.0", | |
| "utmos22_strong", | |
| trust_repo=True, | |
| ) | |
| _PREDICTOR = _PREDICTOR.to(_DEVICE).eval() | |
| return _PREDICTOR | |
| def evaluate_utmos(audio_path: str) -> dict: | |
| if not audio_path or not Path(audio_path).exists(): | |
| raise ValueError("الصوت الناتج غير موجود.") | |
| wav, sr = sf.read(audio_path, dtype="float32", always_2d=False) | |
| wav = np.asarray(wav, dtype=np.float32) | |
| if wav.ndim > 1: | |
| wav = wav.mean(axis=1) | |
| if wav.size == 0: | |
| raise ValueError("الصوت الناتج فارغ.") | |
| predictor = _load() | |
| tensor = torch.from_numpy(wav).unsqueeze(0).to(_DEVICE) | |
| with torch.inference_mode(): | |
| score = float(predictor(tensor, int(sr))[0].detach().cpu().item()) | |
| return { | |
| "metric": "UTMOS Naturalness", | |
| "score": score, | |
| "scale": "1-5", | |
| "higher_is_better": True, | |
| "model": "tarepan/SpeechMOS:v1.2.0 / utmos22_strong", | |
| } | |
| def release_utmos(): | |
| global _PREDICTOR | |
| with _LOCK: | |
| _PREDICTOR = None | |
| gc.collect() | |
| if torch.cuda.is_available(): | |
| torch.cuda.empty_cache() | |