Speech Quality Embeddings for Improved Detection and Classification of Degradations in Speech Signals

TL;DR

This paper introduces a method to generate frame-level speech quality embeddings that cluster by degradation type, improving local degradation detection and classification in speech signals.

Contribution

The work extends SSQA models to produce embeddings that distinguish degradation types using contrastive loss and a partial mix-up strategy, enhancing local quality assessment.

Findings

Embedding clusters correspond to different degradation types.

The approach improves degradation detection accuracy.

Out-of-domain data experiments validate robustness.

Abstract

Automatic subjective speech quality assessment (SSQA) traditionally estimates speech quality on an utterance or system level. While this resolution was adequate for older transmission or synthesis systems that produced speech signals of mediocre quality, modern systems generate high-quality speech with degradations that may occur only locally. With suitable model architectures and regularization losses, SSQA models trained with utterance-level targets can also yield useful local predictions of speech quality. In this work, we extend such models to produce frame-level embeddings that cluster by degradation type. Specifically, we employ a partial mix-up strategy on a parallel corpus of clean and degraded utterances and apply a contrastive loss to distinguish between degradation types. Through experiments on both in- and out-of-domain data, we demonstrate that our approach improves degradation detection and enables the identification of degradation types by analyzing embedding clusters.