Pushing the performances of ASR models on English and Spanish accents

TL;DR

This paper demonstrates that using pre-trained embeddings and auxiliary classification losses can enhance speech-to-text ASR models across different accents and languages, promoting more universal performance.

Contribution

It introduces simple methods like pre-trained embeddings and auxiliary losses to improve ASR performance across multiple models and languages.

Findings

Improved ASR accuracy on English and Spanish accents.

Enhanced model robustness with auxiliary classification losses.

Effective across various model architectures.

Abstract

Speech to text models tend to be trained and evaluated against a single target accent. This is especially true for English for which native speakers from the United States became the main benchmark. In this work, we are going to show how two simple methods: pre-trained embeddings and auxiliary classification losses can improve the performance of ASR systems. We are looking for upgrades as universal as possible and therefore we will explore their impact on several models architectures and several languages.