MFSC系数特征局部有限权重共享CNN语音识别  

Local Finite Weight Sharing of MFSC Coefficients Based CNN Speech Recognition

作　　者：黄玉蕾[1] 罗晓霞[2] 刘笃仁[3] 

机构地区：[1]西安培华学院中兴电信学院,西安710125 [2]西安科技大学计算机科学与技术学院,西安710054 [3]西安电子科技大学电子工程学院,西安710126

出　　处：《控制工程》2017年第7期1507-1513,共7页Control Engineering of China

基　　金：陕西省教育科学十三五规划2016年度规划课题(SGH16H258)

摘　　要：针对传统语音识别应用中识别效果不理想的问题,提出一种基于美尔谱系数(MFSC)特征的有限局部权重共享卷积神经网络(CNN)语音识别。首先,借鉴图像处理中对输入图像信息的处理方式,构建语音信号的二维阵列特征映射输入形式,每个映射表示为含静态数据、一阶导数、二阶导数的MFSC系数特征,便于应用图像处理方式进行识别;其次,引入图像处理的卷积神经网络,并且针对语音信号特征的局部特性,构建有限局部权重共享卷积神经网络学习算法,提高语音信号辨识度并降低算法复杂度;最后,通过实验对所提算法进行验证,并且给出算法参数变化影响实验,为具体应用提供依据。According to the problem of non-ideal identification in traditional speech recognition applications,here the local finite weight sharing of MFSC coefficients based CNN speech recognition is proposed. Firstly,learning from the image processing method, the two-dimensional array feature map input form of the speechsignal is constructed. Each mapping is represented by the MFSC coefficients of the static data, the firstderivative and the second order derivative, which is convenient for the application of the image processingmethod; Secondly, the neural network of image processing is introduced into speech recognition, and accordingto the local characteristics of the speech signal, the finite local weight sharing convolutional neural network isconstructed to improve the speech recognition degree and reduce the complexity of the algorithm; Finally, theproposed algorithm is verified by experiments, and the parameter selection experiment of the algorithm is alsopresented.

关 键 词：卷积神经网络 局部权重共享 语音识别 特征映射 美尔谱系数 

分 类 号：TP912[自动化与计算机技术]