ISCA Archive Interspeech 2023
ISCA Archive Interspeech 2023

Improving End-to-End Modeling For Mandarin-English Code-Switching Using Lightweight Switch-Routing Mixture-of-Experts

Fengyun Tan, Chaofeng Feng, Tao Wei, Shuai Gong, Jinqiang Leng, Wei Chu, Jun Ma, Shaojun Wang, Jing Xiao

Code-switching is a common phenomenon in multilingual communities. In this paper, we study end-to-end model for Mandarin-English intra-sentential code-switching speech recognition. A lightweight Switch-Routing network is proposed, which includes two experts and a switch router. Two experts, representing Mandarin and English learners, implicitly provide language identification information and skillfully use monolingual data to assist code-switching task training, which solves the problem of data sparsity. In addition, our network is a lightweight structure, which makes use of the advantages of Switch Transformer and discards its weakness of increasing model capacity. Finally, we study the effect of using lightweight Switch Routing in different blocks of encoder and decoder. Compared with Bi-Encoder, proposed model has a better performance on the ASRU code-switching test set, and the most important thing is that it requires much less inference time with RTF decreasing by 31.39%.