Self-supervised learning with bi-label masked speech prediction for streaming multi-talker speech recognition

TL;DR

This paper introduces a novel self-supervised learning method with bi-label masked speech prediction tailored for streaming multi-talker speech recognition, significantly improving performance on overlapping speech transcription tasks.

Contribution

It proposes a new SSL training objective that explicitly preserves representations of all speakers in overlapping speech for streaming scenarios.

Findings

Achieves better word error rates on LibriSpeechMix dataset.

Demonstrates effectiveness of bi-label masked speech prediction in multi-talker streaming recognition.

Shows that conventional SSL techniques are insufficient for overlapping speech.

Abstract

Self-supervised learning (SSL), which utilizes the input data itself for representation learning, has achieved state-of-the-art results for various downstream speech tasks. However, most of the previous studies focused on offline single-talker applications, with limited investigations in multi-talker cases, especially for streaming scenarios. In this paper, we investigate SSL for streaming multi-talker speech recognition, which generates transcriptions of overlapping speakers in a streaming fashion. We first observe that conventional SSL techniques do not work well on this task due to the poor representation of overlapping speech. We then propose a novel SSL training objective, referred to as bi-label masked speech prediction, which explicitly preserves representations of all speakers in overlapping speech. We investigate various aspects of the proposed system including data configuration and quantizer selection. The proposed SSL setup achieves substantially better word error rates on the LibriSpeechMix dataset.