A Likelihood Ratio based Domain Adaptation Method for E2E Models
This work addresses domain adaptation for streaming ASR models like voice assistants, offering an incremental improvement by leveraging text data to enhance rare word recognition in new domains.
The paper tackles the problem of adapting end-to-end automatic speech recognition models to new domains, which is challenging due to their reliance on paired audio-text data and computational expense. It introduces a likelihood-ratio based contextual biasing method that improves rare word recognition, achieving a 10% relative improvement in word error rate on out-of-domain datasets without degrading performance on general data.
End-to-end (E2E) automatic speech recognition models like Recurrent Neural Networks Transducer (RNN-T) are becoming a popular choice for streaming ASR applications like voice assistants. While E2E models are very effective at learning representation of the training data they are trained on, their accuracy on unseen domains remains a challenging problem. Additionally, these models require paired audio and text training data, are computationally expensive and are difficult to adapt towards the fast evolving nature of conversational speech. 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 1-best word error rate (WER) and 10% in n-best Oracle WER (n=8) 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.