Dynamic Bayesian Network And Its Application To Speaker Recognition

Dynamic Bayesian Network(DBN) is a new stochastic model extended from probabilistic network, which incorporates time information into static net and has the ability to deal with time-series data. DBN has several notable features like interpretabil-ity. nonlinearity,extensibility. It can easily incorporate new knowledge in and make complete presentation, inference and learning. Up to now, the theory of DBN is not thoroughly systematic yet, and only a little: applications can be found, but its general features and advantages arouse more and more attentions from researches in all kinds of fields, especially in the field of time-series data analysis. My work of this paper is outlined as the following: First, this work studied and implemented those necessary algorithms within the DBN framework, including topology transformation junction tree creation and global probability propagation. I also studied "frontier algorithm" and "interface algorithm" specifically for the inference in DBN, and implemented the forward-backward algorithm. Several cases of learning in terms of parameter arid structure were discussed. The EM algorithm for speech processing was implemented in the fact that the topology are known and the observations are not complete.Second, this paper made a comparison between the topology of DBN and that of HMM. After discussing several kinds of HMMs, I described how to get a DBN from a HMM and how to get a HMM from a DBN, and then, made a general comparison between those algorithms of inference and learning for the DBN and those for the HMM. Experiment results answered that why DBN is suitable for dealing with those large time-series data.Third,this work presented a DBN-based framework for automatic speaker recognition, and discussed how to use DBN to accomplish the task of training and testing in speaker recognition. Experiments results on YOHO corpus were compared to those achieved by Vector Quantization(VQ), Single Gaussian, Mixture Gaussian Models(GMM), and Hidden Markov Models(HMM). The results showed that our framework has a better performance compared to other methods. It also showed that DBN is a promising way to modelize the speakers" variability.Finally, this work presented a framework for information fusion using DBN in terms of raw data level, feature level, and decision level. Especially, I have incorporated pitch information into acoustic features within the DBN framework. Experiment results showed that the performance of this method exceeds that of a simple combination of pitch information and acoustic information.Although I devote myself studying dynamic bayesian network these years, my work is still a beginning with a long long way to go. The following research work could be approximate inference, learning of structure:. DBN based multi-level fusion, multimodal fusion, embedded classifier, etc.This work is supported by National Natural Science Foundation of P.R..China (No.60273059). National High Technology Research & Development Programme (863) of P.R.China (No.200lAA4180). Zhejiang Provincial Natural Science Foundation for Young Scientist of P.R.China (No.RC01058), Zhejiang Provincial Education Office Foundation (20020721), and Zhejiang Provincial Doctoral Subject Foundation (20020335025).