基于MCE准则的语音识别特征线性判别分析

发布时间：2018-04-28 03:27

本文选题：线性判别分析 + 语音识别　；参考：《自动化学报》2014年06期

【摘要】：提出了一种基于最小分类错误(Minimum classification error,MCE)准则的线性判别分析方法(Linear discriminant analysis,LDA),并将其应用到连续语音识别中的特征变换.该方法采用非参数核密度估计方法进行数据概率分布估计;根据得到的概率分布,在最小分类错误准则下,采用基于梯度下降的线性搜索算法求解判别分析变换矩阵.利用判别分析变换矩阵对相邻帧梅尔滤波器组输出拼接的超矢量变换降维,得到时频特征.实验结果表明,与传统的MFCC特征相比,经过本文判别分析提取的时频特征其识别准确率提高了1.41%,相比于HLDA(Heteroscedastic LDA)和近似成对经验正确率准则(Approximate pairwise empirical accuracy criterion,aPEAC)判别分析方法,识别准确率分别提高了1.14%和0.83%.
[Abstract]:In this paper, a linear discriminant analysis method based on minimum classification error minimum classification error classification criterion is proposed and applied to feature transformation in continuous speech recognition. The nonparametric kernel density estimation method is used to estimate the probability distribution of the data and the linear search algorithm based on gradient descent is used to solve the discriminant analysis transformation matrix under the minimum classification error criterion according to the obtained probability distribution. Using the discriminant analysis transform matrix to reduce the dimension of the output splicing of adjacent frame Mel filter banks, the time-frequency features are obtained. The experimental results show that, compared with the traditional MFCC features, the recognition accuracy of the time-frequency features extracted by the discriminant analysis in this paper is improved by 1.41th, compared with the HLDA(Heteroscedastic LDAs and the approximate empirical accuracy criterion, the approximate pairwise empirical accuracy criteria of the discriminant analysis. The recognition accuracy increased by 1.14% and 0.83% respectively.
【作者单位】：解放军信息工程大学信息系统工程学院;
【基金】：国家自然科学基金(61175017)资助~~
【分类号】：TN912.34

【参考文献】