基于混合因子分析的隐马尔可夫模型  

A Hidden Markov Model Based on Mixture of Factor Analysis

作　　者：王新民[1] 姚天任[2] 

机构地区：[1]孝感学院物理系,湖北孝感432100 [2]华中科技大学电信系,武汉430074

出　　处：《计算机工程与应用》2005年第24期50-52,共3页Computer Engineering and Applications

基　　金：湖北省教育厅重点项目基金资助(编号:2002A02004)

摘　　要：经典隐马尔可夫模型用于语音识别存在的两个主要缺陷是“离散状态假设”和“独立分布假设”。前者忽略了语音信号的非平稳性,后者忽略了语音信号的相关性。文章将混合因子分析方法用于语音建模,提出了基于混合因子分析的隐马尔可夫模型框架,并用动态贝叶斯网络形象地表示。该模型框架不仅从理论上解决了上述问题,而且给出许多语音建模的选择。目前广泛使用的统计声学模型均可视为该模型的特例。The “discrete states assumption”and “conditional independent assumption” are two main limitations in standard hidden Markov model for speech recognition.The former ignores time-short stationarity of speech signals and the latter disconsiders the intra-frame correlation between the feature vector elements.This paper investigates the combined traditional hidden Markov modeling technology with mixture of factor analysis.It proposes a hidden Markov model based on mixture of factor analysis（HMM-MFA） and use a dynamic Bayesian networks for this approach.A number of standard models including HMM-DG（Hidden Markov Model Based Diagonal Gaussian distributions）,HMM-FA（Hidden Markov Model Based Factor Analysis） and other currently very popular acoustic models are forms of HMM-MFA with different configurations.

关 键 词：隐马尔可夫模型 混合因子分析 动态贝叶斯网络 

分 类 号：TN912[电子电信—通信与信息系统]