On the Utility of Self-supervised Models for Prosody-related Tasks

TL;DR

Self-supervised speech models demonstrate strong potential for prosody-related tasks, outperforming baselines and effectively encoding prosodic information, as shown through a new evaluation framework and layerwise analysis.

Contribution

Introduces SUPERB-prosody, a new evaluation framework for assessing SSL speech models on prosody tasks, and demonstrates their effectiveness and layerwise contributions.

Findings

13 of 15 SSL models outperformed baselines on prosody tasks

SSL models performed well on prosody reconstruction and prediction pseudo tasks

Layerwise analysis reveals how models encode prosodic information

Abstract

Self-Supervised Learning (SSL) from speech data has produced models that have achieved remarkable performance in many tasks, and that are known to implicitly represent many aspects of information latently present in speech signals. However, relatively little is known about the suitability of such models for prosody-related tasks or the extent to which they encode prosodic information. We present a new evaluation framework, SUPERB-prosody, consisting of three prosody-related downstream tasks and two pseudo tasks. We find that 13 of the 15 SSL models outperformed the baseline on all the prosody-related tasks. We also show good performance on two pseudo tasks: prosody reconstruction and future prosody prediction. We further analyze the layerwise contributions of the SSL models. Overall we conclude that SSL speech models are highly effective for prosody-related tasks.