Extracting Domain Invariant Features by Unsupervised Learning for Robust Automatic Speech Recognition

TL;DR

This paper proposes using unsupervised learning with FHVAEs to extract domain-invariant features, significantly improving robustness of speech recognition systems across different acoustic conditions.

Contribution

It introduces the use of FHVAE to learn domain-invariant features without supervision, addressing domain mismatch in ASR.

Findings

41% WER reduction on Aurora-4

27% WER reduction on CHiME-4

Effective in unseen acoustic conditions

Abstract

The performance of automatic speech recognition (ASR) systems can be significantly compromised by previously unseen conditions, which is typically due to a mismatch between training and testing distributions. In this paper, we address robustness by studying domain invariant features, such that domain information becomes transparent to ASR systems, resolving the mismatch problem. Specifically, we investigate a recent model, called the Factorized Hierarchical Variational Autoencoder (FHVAE). FHVAEs learn to factorize sequence-level and segment-level attributes into different latent variables without supervision. We argue that the set of latent variables that contain segment-level information is our desired domain invariant feature for ASR. Experiments are conducted on Aurora-4 and CHiME-4, which demonstrate 41% and 27% absolute word error rate reductions respectively on mismatched domains.