Expand description
NISQA — non-intrusive speech-quality learned MOS predictor. NeedsWeights.
NISQA (Mittag et al., 2021) is a deep model (CNN + self-attention) trained to predict speech MOS and its sub-dimensions (noisiness, coloration, discontinuity, loudness) from a single degraded clip, no clean reference. As with DNSMOS, the score IS the trained network; there is no principled hand coded formula that yields it.
This function therefore fails closed: without weights it returns
AudioError::BackendUnavailable and never fabricates a MOS. Fabricating a
learned MOS is forbidden by project rule.
Functions§
- nisqa
- Predict a NISQA speech-quality MOS for
signal.