TigreGotico/utmos-onnx

UTMOS22 strong (ONNX)

MOS prediction for synthesized speech — the VoiceMOS Challenge 2022 strong learner. ONNX export of utmos22_strong from tarepan/SpeechMOS, itself derived from sarulab-speech/UTMOS22. Licence: MIT throughout.

Exported for speechonnxmetrics.

Contract

utmos22_strong.onnx — opset 17, 411 MB

• Input wave, shape ['batch', 'time'], float32 — 16 kHz mono, raw [-1, 1]. There is no resampling in the graph; callers must resample.
• Output mos, shape ['batch'] — a single scalar MOS, already rescaled in-graph. No client-side post-processing.

The time axis is dynamic; the model was exported at 4.0 s and verified across 1/3/5/10 s.

Batch > 1 is not verified. The graph carries no attention padding mask, so zero-padded ragged batching would produce wrong scores. Score one clip per call.

Export fidelity

Verified against a fresh torch.hub load of the unpatched model on 21 cases — 12 real clips spanning MOS 1.26–4.50, a duration sweep, and five odd-sample-count inputs. Max absolute MOS deviation: 6.20e-06.

fairseq_alt.pad_to_multiple branches on float.is_integer() over a traced tensor size. That padding is inert — padded frames are masked out of attention and then sliced off — so the export takes the no-pad path. Odd frame counts were tested specifically to exercise the removed branch (max deviation 4.65e-06 there).

export_utmos.py reproduces the export.

Citation

Saeki et al., UTMOS: UTokyo-SaruLab System for VoiceMOS Challenge 2022, Interspeech 2022.