Two-pass Approach for Recognizing Code-Switching Speech

摘要

We propose a recognition method for code-switching utterances in multilingual speech recognition. Code-switching is the concurrent use of more than one language in a single conversation. From a linguistic point of view, the grammatical aspect of code-switching has always been an issue since there is no clear grammatical rules for it. In addition to that, the sparseness of code-switching text corpus makes it more difficult to build a language model for speech recognition. We propose a two-pass language modeling approach that takes knowledge from monolingual language models to make a good representation of the code-switching grammar. We evaluate our proposed method using a WFST-based Indonesian-English ASR system for a computer voice command task and prove its effectiveness.