说话人识别及其应用的研究

虽然理论上隐马尔可夫模型（HMM）是较为有效的一种说话人识别方法，但传统的模型训练方法──Baum－Welch算法不仅运算量和存储量较大，而且若因经验不足、模型初值设置不当会导致算法发散或迭代收敛到非全局最优点。本文提出一种新的方法，将状态分割、动态聚类、模糊统计与传统的Baum－Welch算法相结合应用于说话人识别，既降低了运算量和存储量，又避免了因初值设置不当而导致算法迭代收敛到非全局最优点。本文在大量实验的基础上，建立了说话人识别系统并进行了实验研究，收到了良好的效果。该系统模型数目少，运算复杂度低，可扩充性强，易于训练，便于识别，具有广阔的应用前景。

Application of Text-Independent Speaker Recognition

Since the widely used Hidden Markov Models (HMM) in speaker recognition is difficult to be implemented for high computation complexity and large memory needs. Based on a lot of experiments,a new method is proposed in this paper and a speaker recognition system is built by using state-Deviding, Dynamic Cluster,Fuzzy statistic and traditional Baum - Welch algorithm. Features of the system are less models, low memory computation, low computation complexity, easy of expand ing.