Self-supervised speech quality assessment (S3QA): Leveraging speech foundation models for a scalable speech quality metric

Le résumé fourni par la source

Methods for automatically assessing speech quality in real world environments are critical for robust human language technologies and assistive devices. Behavioral ratings (e.g., mean opinion scores) are considered the gold standard, but are susceptible to inter-rater variability, cannot be generalized, and are labor-intensive, thus limiting the acoustic challenges they can quantify. We present a scalable method for assessing speech quality: the self-supervised speech quality assessment model. First, we manipulated high-quality utterances, using acoustic challenges that emulate real-world degradation: filtering, reverberation, background noise, and digital compression. Second, we leveraged a pre-trained speech foundation model, WavLM, to derive cosine distances between the clean and degraded versions of each utterance in the embedding space. A transformer-based model was trained to predict these cosine distances, given only the degraded utterances. The trained model was evaluated on unseen corpora of synthetic mixtures, NISQA, and VOiCES. The self-supervised speech quality assessment model accurately predicts degradation cosine distances across a wide range of acoustic challenges and is aligned with behavioral ratings (mean opinion scores), automatic speech recognition performance and other important features (microphone distances). The model is available online.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.