Extracting speaker and emotion information from self-supervised speech models via channel-wise correlations

TL;DR

This paper explores a novel correlation pooling method to extract speaker and emotion information from self-supervised speech models, demonstrating improved performance over traditional mean pooling techniques.

Contribution

It introduces correlation pooling as an alternative to mean pooling for aggregating speech representations, showing enhanced extraction of speaker and emotion features.

Findings

Correlation pooling outperforms mean pooling in extracting speaker and emotion info

Fusion of pooling methods yields further performance gains

Code implementation is publicly available for reproducibility

Abstract

Self-supervised learning of speech representations from large amounts of unlabeled data has enabled state-of-the-art results in several speech processing tasks. Aggregating these speech representations across time is typically approached by using descriptive statistics, and in particular, using the first- and second-order statistics of representation coefficients. In this paper, we examine an alternative way of extracting speaker and emotion information from self-supervised trained models, based on the correlations between the coefficients of the representations - correlation pooling. We show improvements over mean pooling and further gains when the pooling methods are combined via fusion. The code is available at github.com/Lamomal/s3prl_correlation.