More Speaking or More Speakers?

TL;DR

This paper analyzes how the number of speakers in labeled and unlabeled datasets affects the performance of SSL and ST methods in speech recognition, revealing different data requirements for each approach.

Contribution

It provides a systematic analysis of the impact of speaker diversity and dataset composition on SSL and ST speech recognition methods, which was previously unexplored.

Findings

SSL needs large unlabeled datasets for high accuracy

ST benefits from sufficient speaker diversity in labeled data

Different data regimes favor different semi-supervised approaches

Abstract

Self-training (ST) and self-supervised learning (SSL) methods have demonstrated strong improvements in automatic speech recognition (ASR). In spite of these advances, to the best of our knowledge, there is no analysis of how the composition of the labelled and unlabelled datasets used in these methods affects the results. In this work we aim to analyse the effect of number of speakers in the training data on a recent SSL algorithm (wav2vec 2.0), and a recent ST algorithm (slimIPL). We perform a systematic analysis on both labeled and unlabeled data by varying the number of speakers while keeping the number of hours fixed and vice versa. Our findings suggest that SSL requires a large amount of unlabeled data to produce high accuracy results, while ST requires a sufficient number of speakers in the labelled data, especially in the low-regime setting. In this manner these two approaches improve supervised learning in different regimes of data composition.