基于隐马尔可夫模型(HMM)的声母识别

A SPEECH RECOGNITION SYSTEM FOR CONSONANTS BASED ON HIDDEN MARKOV MODELS (HMM)

本文研究了在声母分类基础上,建立一个以音节为输入单元,采用隐马尔可夫模型(HMM)识别声母的语音系统;针对声母特征的不稳定性及码字序列的模糊性,提出了码字替代的理论和算法,还提出了HMM的自适应和自学习算法,为建立非特定人的语音识别系统打下基础。目前本系统声母的识别率对特定人是85％左右,对非特定人达78％左右。

We' ve set up a speech recognition system with the input unit of syllables based on the classification of initial consonants. The approach of probabilistic statistics based on Hidden Markov Models (HMM) is used. To address the non-stability of initial consonants and the fuzziness of code words, substitution theory and algorithm are proposed. Meanwhile we propose an algorithm of adaptation of HMM so as to set up a speaker-independent recognition system. Now the recognition rate of initial consonants is about 85% for the specker-dependent system and about 78% for the specker-independent system.