Reducing Domain mismatch in Self-supervised speech pre-training

Abstract

Masked speech modeling (MSM) methods such as wav2vec2 or w2v-BERT learn representations over speech frames which are randomly masked within an utterance. While these methods improve performance of Automatic Speech Recognition (ASR) systems, they have one major limitation. They treat all unsupervised speech samples with equal weight, which hinders learning as not all samples have relevant information to learn meaningful representations. In this work, we address this limitation. We propose ask2mask (ATM), a novel approach to focus on specific samples during MSM pre-training. ATM employs an external ASR model or scorer to weight unsupervised input samples by performing a fine-grained data selection. ATM performs masking over the highly confident input frames as chosen by the scorer. This allows the model to learn meaningful representations. We conduct fine-tuning experiments on two well-benchmarked cor-pora: LibriSpeech (matching the pre-training data) and, AMI and CHiME-6 (not matching the pre-training data). The results substantiate the efficacy of ATM on significantly improving the recognition performance under mismatched conditions while still yielding modest improvements under matched conditions.