基于混合因子分析的隐马尔可夫模型

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。

A Hidden Markov Model Based on Mixture of Factor Analysis

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 MarkovModel Based Factor Analysis) and other currently very popular acoustic models are forms of HMM-MFA with different configurations.