Multilingual Transformer Language Model for Speech Recognition in Low-resource Languages

TL;DR

This paper introduces a locale-group multilingual Transformer language model that improves speech recognition in low-resource languages by enhancing performance and reducing costs, outperforming traditional models and aiding monolingual adaptation.

Contribution

The study proposes a novel locale-grouping approach for multilingual Transformer LMs that boosts ASR performance and lowers maintenance costs in low-resource language settings.

Findings

Locale-group multilingual LMs outperform traditional models.

Fine-tuned locale-group models yield better monolingual LMs.

Significant reduction in operational expenses.

Abstract

It is challenging to train and deploy Transformer LMs for hybrid speech recognition 2nd pass re-ranking in low-resource languages due to (1) data scarcity in low-resource languages, (2) expensive computing costs for training and refreshing 100+ monolingual models, and (3) hosting inefficiency considering sparse traffic. In this study, we present a new way to group multiple low-resource locales together and optimize the performance of Multilingual Transformer LMs in ASR. Our Locale-group Multilingual Transformer LMs outperform traditional multilingual LMs along with reducing maintenance costs and operating expenses. Further, for low-resource but high-traffic locales where deploying monolingual models is feasible, we show that fine-tuning our locale-group multilingual LMs produces better monolingual LM candidates than baseline monolingual LMs.