Code-mixing is a common phenomenon in bilingual societies. It refers to the intra-sentential switching of two different languages in a spoken utterance. This paper presents the first study on automatic recognition of Cantonese-English code-mixing speech, which is common in Hong Kong. This study starts with the design and compilation of code-mixing speech and text corpora. The problems of acoustic modeling, language modeling, and language boundary detection are investigated. Subsequently, a large-vocabulary code-mixing speech recognition system is developed based on a two-pass decoding algorithm. For acoustic modeling, it is shown that cross-lingual acoustic models are more appropriate than language-dependent models. The language models being used are character tri-grams, in which the embedded English words are grouped into a small number
of classes. Language boundary detection is done either by exploiting the
phonological and lexical differences between the two languages or is done based on the result of cross-lingual speech recognition. The language boundary information is used to re-score the hypothesized syllables or words in the decoding process. The proposed code-mixing speech recognition system attains the accuracies of 56.4% and 53.0% for the Cantonese syllables and English words in code-mixing utterances.