Integrating Text Inputs For Training and Adapting RNN Transducer ASR Models

TL;DR

This paper introduces a new text-based training method for RNN Transducer ASR models that enables effective domain adaptation using only text data, significantly reducing word error rates across multiple datasets.

Contribution

It presents a novel text representation and training framework that allows RNN-T models to adapt their internal language model component with only text data, improving flexibility and customization.

Findings

13% WER reduction on Switchboard and CallHome datasets

20-45% relative WER reduction in domain-specific adaptations

Effective domain adaptation using only unpaired text data

Abstract

Compared to hybrid automatic speech recognition (ASR) systems that use a modular architecture in which each component can be independently adapted to a new domain, recent end-to-end (E2E) ASR system are harder to customize due to their all-neural monolithic construction. In this paper, we propose a novel text representation and training framework for E2E ASR models. With this approach, we show that a trained RNN Transducer (RNN-T) model's internal LM component can be effectively adapted with text-only data. An RNN-T model trained using both speech and text inputs improves over a baseline model trained on just speech with close to 13% word error rate (WER) reduction on the Switchboard and CallHome test sets of the NIST Hub5 2000 evaluation. The usefulness of the proposed approach is further demonstrated by customizing this general purpose RNN-T model to three separate datasets. We observe 20-45% relative word error rate (WER) reduction in these settings with this novel LM style customization technique using only unpaired text data from the new domains.