Boosting Cross-Domain Speech Recognition with Self-Supervision

The cross-domain performance of automatic speech recognition (ASR) could be severely hampered due to the mismatch between training and testing distributions. Since target domain usually lacks labeled data, shifts exist at acoustic linguistic levels, it is challenging perform unsupervised adaptation (UDA) for ASR. Previous work has shown that self-supervised learning (SSL) or pseudo-labeling (PL) effective in UDA by exploiting self-supervisions unlabeled data. However, these also face degradation mismatched distributions, which previous fails address. This presents a systematic framework fully utilize data with self-supervision pre-training fine-tuning paradigm. On one hand, we apply continued replay techniques mitigate SSL pre-trained model. other propose domain-adaptive approach based on PL technique three unique modifications: Firstly, design dual-branch method decrease sensitivity erroneous pseudo-labels; Secondly, devise an uncertainty-aware confidence filtering strategy improve pseudo-label correctness; Thirdly, introduce two-step incorporate knowledge, thus generating more accurate pseudo-labels. Experimental results various scenarios demonstrate proposed effectively boosts significantly outperforms approaches.