Fisher算法在语音声学模型分析中的研究与应用

摘要

语音识别中，通常把发音过程看作短时平稳的随机过程，为同时兼顾到平稳性和短时性，马尔可夫链成为语音建模的有效工具。本文在声学模型研究中，首次引进了多元统计分析的理论，用Fisher算法对语音模型的状态中心分类进行了研究，提出了一种新的基于Fisher算法的状态中心估计方法，并同时指出了在汉语语音识别中，HMM状态数宜取在6～8之间。%In the sound recognition, the sound process is usually viewed as temporal and steady randomization. In order to get steadiness and temporality simultaneously, the "Malcolf Chain" becomes the effect tool of building sound models. This paper firstly contributes the theory of pluralistic statistic analysis to the studies on acoustic models, then studies the classification of state method and estimates the state center based on Fisher. Meanwhile, it brings forward that the HMM state number should be better chosen between 6 and 8 in Chinese sound recognition.