A likelihood ratio based domain adaptation method for E2E models

End-to-end (E2E) automatic speech recognition models like Recurrent Neural Networks Transducer (RNN-T) are becoming a popular choice for streaming applications like voice assistants. While E2E models are very effective at learning representation of the training data they are trained on, adapting them to unseen domains remains a challenging problem, as these models typically require large amounts of training data. Additionally, training these models is computationally expensive and are difficult to adapt towards the fast evolving nature of conversational speech, common in voice assistants. In this work, we explore a contextual biasing approach using likelihood-ratio that leverages text data sources to adapt RNN-T model to new domains and entities. We show that this method is effective in improving rare words recognition, and results in a relative improvement of 10% in word error rate (WER) and 10% in Oracle WER on multiple out-of-domain datasets without any degradation on a general dataset. We also show that complementing the contextual biasing adaptation with adaptation of a second-pass rescoring model gives additive WER improvements.