Expand description
DNSMOS — non-intrusive (no-reference) learned MOS predictor. NeedsWeights.
DNSMOS (Microsoft, ICASSP 2021) is a learned model: a CNN trained on crowd-sourced ratings predicts a Mean Opinion Score from a single degraded clip, with no clean reference. There is no closed-form, hand-written way to produce that number — the score IS the network. Without the trained weights we cannot compute a MOS, and the project rule is absolute: never fabricate a MOS.
Therefore this function fails closed: with no weights it returns
AudioError::BackendUnavailable. It does not, under any circumstance, return
an invented number. When weights are supplied the loader is likewise not yet
present, so it also fails closed rather than guessing — surfaced as
BackendUnavailable (backend not wired), not as a fake score.
Functions§
- dnsmos
- Predict a DNSMOS quality score for
signal.