一种基于径向基函数神经网络的语音降噪方法

A METHOD TO DECREASE NOISE IN SPEECH BASED ON RBF NEURAL NETWORK

采用径向基函数神经网络在时域上对含噪语音信号进行降噪处理.针对语音信号的短时平稳性以及噪声的随机性,对语音信号进行分帧预处理;用分帧后的纯净语音信号作为径向基函数网络的教师信号,并利用Matlab神经网络工具箱设计和训练网络.实验结果表明,径向基函数网络作为语音信号滤波器,可有效地抑制语音信号中的白噪声,具有良好的降噪性能.

In this paper, the RBF neural network is adopted to process the denoising speech signal for decreasing noise. As a result of speech signal＇s short-time stationarity and the randomicity of noise, the pre-process of enframing the speech signal is carried out firstly. Then the pure speech signal required by enframing is employed as the teacher signal of RBF neural network. Thus the network is designed and trained on the platform of the MATLAB toolbox of neural network. According to the result of experiment, as the filter of speech signal, RBF neural network can effectively restrain the white noise. By this means the capability of decreasing noise is greatly improved.