Third International Conference on Spoken Language Processing (ICSLP 94)
Presently many speaker recognition algorithms can provide high accuracy. They need reliable samples of the parameter space, long utterances and large amount of training data for each speaker. In practical applications, data acquisition constraints thus limit their domain of application. This paper proposes a speaker recognition method that creates models to specify speaker information accurately by using only a small amount of training data and very short utterance for each speaker, both for training and recognition. The proposed methods based on a discrete hidden Markov model(HMM) improves modeling of output probability estimation. The basic idea is that the proposed hidden Markov VQ model(HMVQM) uses the state dependent codebook, and each state represents a partition of speaker information. This method can be considered as multisection codebook model with stochastic transitions between section. Speaker identification experiments based on single syllable word tests give a 1.5% error rate for the proposed HMVQM method, whereas the discrete HMM method gives error rate of 24.12%. In this experiment, We use only two training data for each speaker.
Bibliographic reference. Yun, Seong Jin / Oh, Yung Hwan (1994): "Performance improvement of speaker recognition system for small training data", In ICSLP-1994, 1863-1866.