一种基于图信号处理的BP神经网络语音识别方案

A BP neural network speech recognition scheme based on graph signal processing

文中研究图信号处理技术在语音信号处理领域的实现和应用。基于语音信号基本特征,提出一种语音图信号拓扑结构和图邻接矩阵新方案,即语音采样点为图顶点,全连接语音拓扑结构和语音样值差递减幂函数作为边权重。分析了新方案下浊音、清音和静音的图傅里叶变换。基于新方案下语音图傅里叶变换系数集中于低频的特性,设计图低通滤波器,并将该低通滤波器滤波后的语音信号进行基于BP神经网络的数字语音识别实验,正确识别率获得明显提高。

This paper applies the graph signal processing technology to speech signal processing.Based on the basic characteristics of speech signals,a new method of topology structure and graph adjacency matrix of speech graph signals is proposed.This method uses speech sampling points as graph vertices,establishes the fully⁃connected speech topology,and assigns decreasing power function of value difference of samples as the edge weights.Then,the graph Fourier transform of voiced,unvoiced and non⁃speech signals are analyzed by the proposed approach.Since the graph Fourier transform coefficients are accumulated at the low frequency,a low pass graph filter is designed.The recognition performance of speech signals after filtering is increased significantly based on the BP neural network.