Distilling the Knowledge of BERT for Sequence-to-Sequence ASR
This work addresses the problem of context access in ASR for speech recognition applications, offering an incremental improvement by integrating BERT without extra inference cost.
The paper tackles the limitation of left-to-right decoding in sequence-to-sequence ASR by using BERT for knowledge distillation to incorporate both left and right context, resulting in significant performance improvements on the Corpus of Spontaneous Japanese dataset compared to baselines.
Attention-based sequence-to-sequence (seq2seq) models have achieved promising results in automatic speech recognition (ASR). However, as these models decode in a left-to-right way, they do not have access to context on the right. We leverage both left and right context by applying BERT as an external language model to seq2seq ASR through knowledge distillation. In our proposed method, BERT generates soft labels to guide the training of seq2seq ASR. Furthermore, we leverage context beyond the current utterance as input to BERT. Experimental evaluations show that our method significantly improves the ASR performance from the seq2seq baseline on the Corpus of Spontaneous Japanese (CSJ). Knowledge distillation from BERT outperforms that from a transformer LM that only looks at left context. We also show the effectiveness of leveraging context beyond the current utterance. Our method outperforms other LM application approaches such as n-best rescoring and shallow fusion, while it does not require extra inference cost.