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

  • 2024-07-09 13:14:48
  • Mengjie Qian, Siyuan Tang, Rao Ma, Kate M. Knill, Mark J. F. Gales
  • 0


Foundation ASR models often support many languages, e.g. 100 languages inWhisper. However, there has been limited work on integrating an additional,typically low-resource, language, while maintaining performance on the originallanguage set. Fine-tuning, while simple, may degrade the accuracy of theoriginal 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 areoptimised. Elastic Weight Consolidation (EWC) offers an alternative compromisewith the potential to maintain performance in specific target languages.Results show that direct fine-tuning yields the best performance for the newlanguage but degrades existing language capabilities. EWC can address thisissue for specific languages. If only adaptation parameters are used, thelanguage capabilities are maintained but at the cost of performance in the newlanguage.


Quick Read (beta)

loading the full paper ...