DNSMOS (Deep Noise Suppression Mean Opinion Score)

DNSMOS is a neural network trained on human listening-test data to predict mean opinion scores for speech recordings. It was developed for the Deep Noise Suppression Challenge and reports an overall score alongside sub-scores for speech quality and for background noise. A personalized variant judges quality relative to a target speaker.

Because it needs no clean reference signal, DNSMOS can be run on any recording, including material for which no ground truth exists. That same property is its main limitation: it rates what it hears, so a track can score well while containing the wrong speaker's words. It is normally reported alongside reference-based measures rather than on its own.