基于群稀疏约束的语音识别特征混合判别分析  

Group-Lasso Based Mixture Discriminant Analysis Method for Speech Recognition Feature

作　　者：陈斌[1] 陈琦[1] 张连海[1] 屈丹[1] 李弼程[1] 

机构地区：[1]解放军信息工程大学信息系统工程学院,河南郑州450002

出　　处：《四川大学学报（工程科学版）》2015年第5期139-145,共7页Journal of Sichuan University (Engineering Science Edition)

基　　金：国家自然科学基金资助项目(61175017;61403415)

摘　　要：为了克服因数据不足而造成较难提取稳定的长时特征的问题,提出一种基于群稀疏约束的混合判别分析方法。该方法首先采用高斯混合模型描述数据的分布,在此基础上利用2次变分的形式进行群稀疏的表示,得到基于群稀疏约束的混合判别分析目标函数。接着,通过定义模糊响应矩阵(blurred response matrix),有效地结合最优化得分方法求解判别分析变换矩阵。最后,拼接相邻帧梅尔滤波器组输出组成超矢量,采用变换矩阵进行变换降维,提取时频特征。实验结果表明,提出的方法能够得到稀疏的变换矩阵,相比于PLDA(penalized LDA)和SLDA(sparse LDA)判别分析方法,识别准确率分别提高了0.71%和1.53%,且在数据不足的条件下,该方法能获得更高的识别性能。In order to extract the stable long-term features when the data is insufficient,a group-Lasso based mixture discriminant analysis method was proposed. Firstly,the Gaussian mixture model was used to describe the distribution of data,and the objective function of group-Lasso based mixture diseriminant analysis was got based on the quadratic variational form of the group-Lasso. Subsequently,through defining the blurred response matrix,the problem of solving the discriminant analysis transform matrix was figured out by effectively combined with the optimal scoring method. Finally,the super-vector was obtained by conjoined the adja- cent frames Mel filter bank output,and the time-frequency feature was extracted after the dimensionality of super-vector reduced using the transform matrix. The experimental results showed that this method can obtain sparse transformation matrix. Compared to the PLDA and SLDA discriminant analysis method,the recognition accuracy rate is increased by 0.71% and 1.53% respectively, and when lack of data, this method can achieve higher recognition performance.

关 键 词：混合判别分析 群稀疏 特征变换 语音识别 

分 类 号：TN912.3[电子电信—通信与信息系统]