基于连续高斯混合密度HMM的汉语全音节语音识别研究

A study of speech recognition of complete Chinese syllables based on continuous Gaussian Mixture HMM

本文在大量语音分析实验的基础上，对ＨＭＭ用于汉语全音节语音识别进行了较为深入的探讨，建立了一个连续高斯混合密度ＨＭＭ的汉语全音节语音识别系统．该系统在训练算法上撇开了传统的Ｂａｕｍ－Ｗｅｌｃｈ算法，代之以计算复杂度小、存储量小、迭代次数少且具有自动分割效应的分段Ｋ平均算法。对于ＨＭＭ的模型单元的选择，单元的结构以及模型参数的选取，充分考虑了汉语语音的特点；并在语音特征上做了深入的实验分析工作，采用了符合人耳听觉特性的Ｍｅｌ－Ｓｃａｌｅｄ参数，用ＦＦＴ倒谱代替了ＬＰＣ倒谱，同时利用了语音的动态谱特征和能量特征。另外，本文还针对汉语声母的特点，独特地提出了变帧移分析策略。整个识别系统的首选正识率为９１．１％．

Based on a large amount of speech study and experiments,this paper gives a deep study on how HMM is applied to the Chinese speech recognition,and establishal a speech recognition system of complete Chinese syllables using the continuous Gaussian Mixture HMM. The systems does not adopt the traditional Baum-Welch Algorithm, but uses segmental K-Means Training.which needs much smaller memory,calculation and iteration times,and can give automatic segmentation of Speech.On the choise of HMM unit,unit structure,and unit parameters,the poper gives a thorough consideration for the properties of Chinese speech.The paper also gives a deep study on speech features,and employed Mel-Scaled FFT-CEP (instead of LPC-CEP) and its regression coefficients,normalized log-energy and its regression coefficients.In addition,the paper proposes the Variant Frame Shift Analysis Algorithm considering characteristics of consonants.The system recognition rate is 91.1 %.