Large-vocabulary speech recognition system for Taiwanese (Min-nan)

In this paper an initial study and some preliminary work about Taiwanese (Min-nan, Southern Hokkian) speech recognition has been described, including a set of phonetic transcription symbols, a Taiwanese pronunciation lexicon of more than 50 thousand words, several sets of phonetically balanced words, and a set of speech data. The inter-syllabic right context dependent Initial/Finals or phonemes are shown to be very useful in the acoustic modeling. Furthermore, we adopted not only model clustering based on an acoustic decision tree to improve data sharing, but also a hybrid duration model to improve the accuracy of state duration modeling in CHMM. Promising recognition rates and a satisfactory recognition speed can be achieved. A prototype real-time speech recognition system has been constructed and some preliminary experimental results have also been reported. The recognition task defined here is a large-vocabulary, isolated-word (multi-syllabic), and speaker-dependent system.