Abstract

Multilingual speech recognition with neural networks is often implementedwith batch-learning, when all of the languages are available before training.An ability to add new languages after the prior training sessions can beeconomically beneficial, but the main challenge is catastrophic forgetting. Inthis work, we combine the qualities of weight factorization, transfer learningand Elastic Weight Consolidation in order to counter catastrophic forgettingand facilitate learning new languages quickly. Such combination allowed us toeliminate catastrophic forgetting while still achieving performance for the newlanguages comparable with having all languages at once, in experiments oflearning from an initial 10 languages to achieve 27 languages

 

Quick Read (beta)