HMM语音识别模型与一种修正训练算法  

A Study on Hidden Markov Models Used in Speech Recognition and a Modified Training Algorithm

作　　者：关存太 陈永彬[1] 吴伯修 

机构地区：[1]东南大学无线电工程系

出　　处：《东南大学学报（自然科学版）》1994年第1期17-24,共8页Journal of Southeast University：Natural Science Edition

基　　金：国家自然科学基金

摘　　要：本文在统一的框架下描述了隐马尔柯夫模型（ＨＭＭ）用于语音识别时的各种形式，包括离散ＨＭＭ、连续混合密度ＨＭＭ、半连续ＨＭＭ和最大分量连续ＨＭＭ等，指出各种模型均是统一形式下的导出形式．文中就离散ＨＭＭ、连续混合密度ＨＭＭ和最大分量连续ＨＭＭ在非特定人全音节汉语语音识别中的应用，从识别率和复杂度两方面进行了性能比较．为提高最大分量连续ＨＭＭ的识别性能；提出了一种修正的训练算法．Various Hidden Markov Models used in speech recognition in a unified framework aredescribed,These models include discrete HMM,continuous HMM,semi-continuous HMM,andmaximum-component HMM. All these models are derivative forms of the generalized form.Therecognition performances and computational complexities of the discrete HMM,continuous HMM,andmaximum-component HMM are compared with experiments in the task of speaker-independent Chinesespeech recognition with whole syllables.In order to improve the recognition performance of the maximum-component HMM,a modified corrective training algorithm is proposed.

关 键 词：算法 隐马氏模型 元音识别 

分 类 号：TN912.34[电子电信—通信与信息系统]