A Likelihood Ratio based Domain Adaptation Method for E2E Models

TL;DR

This paper presents a likelihood-ratio based domain adaptation method for end-to-end speech recognition models, improving recognition of rare words and out-of-domain data without degrading general performance.

Contribution

It introduces a novel likelihood-ratio based contextual biasing approach that leverages text data for effective domain adaptation of RNN-T models.

Findings

10% relative improvement in 1-best WER on out-of-domain datasets

10% relative improvement in n-best Oracle WER (n=8)

No degradation on general dataset performance

Abstract

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.