Improving ASR for Code-Switched Speech in Under-Resourced Languages Using Out-of-Domain Data

Abstract

We explore the use of out-of-domain monolingual data for the improvement of automatic speech recognition (ASR) of code-switched speech. This is relevant because annotated code-switched speech data is both scarce and very hard to produce, especially when the languages concerned are under-resourced, while monolingual corpora are generally better-resourced. We perform experiments using a recently-introduced small five-language corpus of code-switched South African soap opera speech. We consider specifically whether ASR of English– isiZulu code-switched speech can be improved by incorporating monolingual data from unrelated but larger corpora. TDNN-BLSTM acoustic models are trained using various configura-tions of training data. The utility of artificially-generated bilingual English–isiZulu text to augment language model training data is also explored. We find that English-isiZulu speech recognition accuracy can be improved by incorporating mono-lingual out-of-domain data despite the differences between the soap-opera and monolingual speech.