Learn and Don't Forget: Adding a New Language to ASR Foundation Models

TL;DR

This paper explores methods for adding new languages to existing ASR models without degrading their performance on original languages, comparing fine-tuning, adaptation parameters, and EWC techniques.

Contribution

It systematically evaluates adaptation strategies for multilingual ASR models, highlighting trade-offs between performance on new and existing languages.

Findings

Fine-tuning yields best new language performance but degrades original languages.

EWC helps preserve original language performance when adding new languages.

Using only adaptation parameters maintains original language capabilities with some performance loss on the new language.

Abstract

Foundation ASR models often support many languages, e.g. 100 languages in Whisper. However, there has been limited work on integrating an additional, typically low-resource, language, while maintaining performance on the original language set. Fine-tuning, while simple, may degrade the accuracy of the original set. We compare three approaches that exploit adaptation parameters: soft language code tuning, train only the language code; soft prompt tuning, train prepended tokens; and LoRA where a small set of additional parameters are optimised. Elastic Weight Consolidation (EWC) offers an alternative compromise with the potential to maintain performance in specific target languages. Results show that direct fine-tuning yields the best performance for the new language but degrades existing language capabilities. EWC can address this issue for specific languages. If only adaptation parameters are used, the language capabilities are maintained but at the cost of performance in the new language.