| """Automatic naturalness score (UTMOS22 strong, 1-5) so data or recipe changes that |
| hurt audio quality (e.g. 16 kHz training audio) show up next to CER and similarity. |
| Language-agnostic acoustic quality, not a substitute for Nepali listeners.""" |
| import os |
|
|
| _model = None |
| DEV = os.environ.get("EVAL_DEVICE", "cuda") |
|
|
|
|
| def utmos(wav16k): |
| """wav16k: 1-D float numpy array at 16 kHz -> predicted MOS.""" |
| global _model |
| import torch |
| if _model is None: |
| _model = torch.hub.load("tarepan/SpeechMOS:v1.2.0", "utmos22_strong", trust_repo=True).to(DEV).eval() |
| with torch.no_grad(): |
| return float(_model(torch.tensor(wav16k).unsqueeze(0).to(DEV), 16000)) |
|
|
