On-the-Fly Text Retrieval for end-to-end ASR Adaptation

retrieval, language model, domain adaptation, end-to-end ASR, RNN transducer, contextual biasing

End-to-end speech recognition models are improved by incorporat- ing external text sources, typically by fusion with an external lan- guage model. Such language models have to be retrained whenever the corpus of interest changes. Furthermore, since they store the entire corpus in their parameters, rare words can be challenging to recall. In this work, we propose augmenting a transducer-based ASR model with a retrieval language model, which directly retrieves from an external text corpus plausible completions for a partial ASR hy- pothesis. These completions are then integrated into subsequent pre- dictions by an adapter, which is trained once, so that the corpus of interest can be switched without incurring the computational over- head of retraining. Our experiments show that the proposed model significantly improves the performance of a transducer baseline on a pair of question-answering datasets. Further, it outperforms shallow fusion on recognition of named entities by about 7% relative; when the two are combined, the relative improvement increases to 13%