Modular Hybrid Autoregressive Transducer

TL;DR

This paper introduces a modular hybrid autoregressive transducer (MHAT) with separated label and blank decoders, enabling effective text adaptation and improved speech recognition performance.

Contribution

The paper proposes a novel MHAT architecture with separate decoders and an internal LM training method, enhancing text adaptation and LM fusion in end-to-end speech recognition.

Findings

Achieves up to 12.4% relative WER reduction without LM fusion.

Achieves up to 21.5% relative WER reduction with LM fusion.

Effective text adaptation improves recognition accuracy on large-scale data.

Abstract

Text-only adaptation of a transducer model remains challenging for end-to-end speech recognition since the transducer has no clearly separated acoustic model (AM), language model (LM) or blank model. In this work, we propose a modular hybrid autoregressive transducer (MHAT) that has structurally separated label and blank decoders to predict label and blank distributions, respectively, along with a shared acoustic encoder. The encoder and label decoder outputs are directly projected to AM and internal LM scores and then added to compute label posteriors. We train MHAT with an internal LM loss and a HAT loss to ensure that its internal LM becomes a standalone neural LM that can be effectively adapted to text. Moreover, text adaptation of MHAT fosters a much better LM fusion than internal LM subtraction-based methods. On Google's large-scale production data, a multi-domain MHAT adapted with 100B sentences achieves relative WER reductions of up to 12.4% without LM fusion and 21.5% with LM fusion from 400K-hour trained HAT.