Japanese-English Code-Switching Speech Data Construction

Abstract

As the number of Japanese-English bilingual speakers continues to increase, code-switching phenomena also happen more frequently. The units and locations of switches may vary widely from single word switches to whole phrases (beyond the length of the loanword units). Therefore, speech recognition systems must be developed that can handle not only Japanese or English but also Japanese-English code-switching. Consequently, a large-scale code-switching speech database is required for model training. But collecting natural conversation dialogues of Japanese-English data is both time-consuming and expensive. This paper presents the construction of Japanese-English code-switching speech data by utilizing a Japanese and English text-to-speech system from a bilingual speaker. Various switching units are also investigated including units of words and phrases. As a result, we successfully constructed over 280-k speech utterances of Japanese-English code-switching.