Match to Win: Analysing Sequences Lengths for Efficient Self-supervised Learning in Speech and Audio

TL;DR

This paper empirically studies how truncating input sequence lengths during self-supervised learning in speech and audio can significantly reduce computational costs while maintaining performance, enabling more efficient and personalized applications.

Contribution

It provides the first empirical analysis of sequence length effects in SSL pre-training for speech and audio, highlighting the benefits of training on shorter sequences.

Findings

Short sequences reduce resource costs substantially.

Performance remains satisfactory with truncated sequences.

Facilitates SSL training on edge devices.

Abstract

Self-supervised learning (SSL) has proven vital in speech and audio-related applications. The paradigm trains a general model on unlabeled data that can later be used to solve specific downstream tasks. This type of model is costly to train as it requires manipulating long input sequences that can only be handled by powerful centralised servers. Surprisingly, despite many attempts to increase training efficiency through model compression, the effects of truncating input sequence lengths to reduce computation have not been studied. In this paper, we provide the first empirical study of SSL pre-training for different specified sequence lengths and link this to various downstream tasks. We find that training on short sequences can dramatically reduce resource costs while retaining a satisfactory performance for all tasks. This simple one-line change would promote the migration of SSL training from data centres to user-end edge devices for more realistic and personalised applications.