Sub-syllable Acoustic Modeling for Cantonese Speech Recognition

This paper presents a pioneer study on acoustic modeling for continuous Cantonese speech recognition. It starts from the context-independent modeling of sub-syllabic units, namely INITIALs and FINALs, and then moves on to examine a number of context-dependent models that characterize intra-syllable co-articulation. The acoustic models are trained with a large database of Cantonese polysyllabic words and evaluated with a general syllable recognition task in which no lexical or grammatical constraints are incorporated. A syllable recognition accuracy of 67.68% is attained using continuous-density HMM with 4 Gaussian mixtures.