基于多特征组合优化的汉语数字语音识别研究  

Study of Chinese Digital Speech Recognition Based on Various Features Combinatorial Optimization

作　　者：蔡敏[1] 

机构地区：[1]苏州工业园区工业技术学院机电中心,江苏苏州215123

出　　处：《电子器件》2013年第2期282-284,共3页Chinese Journal of Electron Devices

摘　　要：研究了一种汉语数字语音识别方案,首先提取汉语数字语音线性预测倒谱系数(LPCC)和梅尔频率倒谱系数(MFCC)及其一阶差分,并组合成新特征。通过求取其系数矩阵的均值和方差的方式进行一次降维,然后采用基于关联规则的特征选择算法进行二次降维,并采用C4.5决策树算法进行识别。通过实验表明提出的方法能够有效降低特征维度,去除了无用的冗余信息,提高了语音识别率。An Chinese digital speech scheme is studied. Firstly, Linear Prediction Cepstrum Coefficients（LPCC） and Mel Frequency Cepstrum Coefficients（MFCC） and its first differential of the Chinese digital speech are extracted, and composed as new feature set. The means and variances of the coefficient matrix are solved to achieve the first di- mension reduction. To conduct a second dimension reduction, the feature selection approach based on association rule is applied. Then speech recognition is made using decision tree of C4.5 algorithm. The experiment results show that the method proposed in this paper can reduce the dimensions of the feature, wipe off the useless redundant in- formation and improve the recognition rate.

关 键 词：语音识别 线性预测倒谱系数 梅尔频率倒谱系数 特征选择 降维 

分 类 号：H017[语言文字—语言学]