Abstract:

Probability distribution of monophones and triphones in Lhasa Tibetan are calculated and the necessity of establishing a contextual acoustic model in ASR for Lhasa Tibetan is analyzed. Phoneme is chosen as basic unit for acoustic models. According to the characteristics of Tibetan, a pronunciation dictionary based on single syllable is established. Main issues and algorithms for triphone models based on decision tree are discussed. According to IPAs and characteristics of Lhasa dialect, 38 phoneme subsets and question sets for triphone modeling are established. 8170 sentences of 20 speakers are recorded to train the models. Contextual continuous Hidden Markov Models(HMM) based on triphones are established and trained on HTK platform. The recognition results under different sates number and mixtures are analyzed. And the framework for largevocabulary continuous speech recognition of Lhasa Dialect is established.

Key words: Tibetan;Lhasa dialect;LVCSR;HMM;triphone model