Mandarin-English Code-switching Speech Recognition

This work presents the development of a Mandarin-English code-switching speech recognition system. We demonstrate three key novelties in our system. First, we increase our lexicon coverage to 360K words, where phone sets of different languages are maintained separately. Secondly, we used over 1000 hours of training data combining both mono-lingual and code-switch corpus to develop the acoustic model. Finally, for language modelling, we applied context-aware text normalization and word-class language model. When testing on our internal code-switch close talk microphone recording, the system achieves recognition performance that can support real applications.