speechonnxmetrics models
Collection
ONNX weights for the speech-quality metrics used by speechonnxmetrics. • 4 items • Updated
SIGMOS (ONNX)
Multi-dimensional non-intrusive speech quality prediction from Microsoft's Speech Signal Improvement Challenge. Mirrored unmodified for use by speechonnxmetrics.
Upstream: microsoft/SIG-Challenge.
Licence: MIT (see LICENSE.txt).
Contract
model-sigmos_1697718653_41d092e8-epo-200.onnx
- Input
[batch, 3, time, 481]— compressed STFT at 48 kHz - Output 7 sub-scores, in order: COL, DISC, LOUD, NOISE, REVERB, SIG, OVRL
sigmos.py is the upstream reference frontend, kept alongside so the compressed-STFT
preprocessing can be reproduced exactly.
Citation
Ristea et al., ICASSP 2024 Speech Signal Improvement Challenge.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support