Towards Truly Multilingual ASR: Generalizing Code-Switching ASR to Unseen Language Pairs

The paper investigates the generalization of code-switching automatic speech recognition (CS-ASR) to unseen language pairs by employing model merging and domain generalization techniques. The findings indicate that merged bilingual CS-ASR models can modestly transfer capabilities to new language pairs, highlighting a potential pathway for scaling CS-ASR without the need for extensive bilingual datasets for every language combination. This advancement is significant for practitioners as it suggests methods to enhance multilingual ASR systems with limited resources.