Wav2vec‐MoE: An unsupervised pre‐training and adaptation method for multi‐accent ASR

In real life, either the subjective factors of speakers or objective environment degrades performance automatic speech recognition (ASR). This study focuses on one factors, accented speech, and attempts to build a multi-accent ASR system solve degradation caused by different accents, whose characteristic is low resource. To deal with challenge low-resource data styles, wav2vec-MoE (mixture experts) proposed adapt wav2vec 2.0 for ASR. wav2vec-MoE, domain MoE developed introducing pseudo-domain information in pre-training stage, where denotes collection varied same influence factors. The trained two strategies according mismatch assessment between unlabeled target without requiring any explicit information. Experiments show that achieves 14.69% relative word error rate reduction (WERR) AESRC2020 accent dataset an 8.79% WERR Common Voice English dataset.