HYBRID TEXT-INDEPENDENT SPEAKER RECOGNITION USING CHARACTER-BASED BACKGROUND HMMS AND GMMS FOR MANDARIN SPEECH

In mandarin, the words are composed by the concatenation of Chinese characters. In this paper, we propose a hybrid speaker recognition system based on character-based background HMMs and Gaussian mixture models to combine the advantage of them for text-independent Mandarin speech. Here all characters, spoken by all reference speakers selected to form the background HMMs, is represented by a large HMM, named general-character HMM. The estimating process of background model is much easier and simpler than those word or sub-word based HMMs The trained character-based HMMs are used to remove the segments only containing silence and noise from utterances, then the speech segments are used to train the GMMs for text-independent speaker recognition and to specify scoring segments for test utterances. Furthermore, it provides speaker-independent background likelihood scores for verification. The normalization effects using the background HMMs with different topological structures are compared. It is shown that score normalization using the background model can improve the verification performance greatly, but the topological structure of general character HMM for Mandarin speech should be defined appropriately. 1