Cross-Modal Transformer-Based Neural Correction Models for Automatic Speech Recognition

TL;DR

This paper introduces a novel cross-modal transformer-based neural correction model that refines ASR outputs by jointly encoding speech and text inputs using cross-modal self-attention, improving accuracy over traditional methods.

Contribution

The paper presents a new cross-modal transformer model that jointly encodes speech and text inputs for neural correction, capturing their relationships more effectively.

Findings

Achieved better ASR performance than conventional neural correction models.

Demonstrated effectiveness on Japanese natural language ASR tasks.

Utilized cross-modal self-attention to improve error correction.

Abstract

We propose a cross-modal transformer-based neural correction models that refines the output of an automatic speech recognition (ASR) system so as to exclude ASR errors. Generally, neural correction models are composed of encoder-decoder networks, which can directly model sequence-to-sequence mapping problems. The most successful method is to use both input speech and its ASR output text as the input contexts for the encoder-decoder networks. However, the conventional method cannot take into account the relationships between these two different modal inputs because the input contexts are separately encoded for each modal. To effectively leverage the correlated information between the two different modal inputs, our proposed models encode two different contexts jointly on the basis of cross-modal self-attention using a transformer. We expect that cross-modal self-attention can effectively capture the relationships between two different modals for refining ASR hypotheses. We also introduce a shallow fusion technique to efficiently integrate the first-pass ASR model and our proposed neural correction model. Experiments on Japanese natural language ASR tasks demonstrated that our proposed models achieve better ASR performance than conventional neural correction models.