基于改进压缩感知的说话人识别抗噪算法

摘要

压缩感知CS（compressive sensing）是一种基于信号稀疏性，有效提取信号中有用信息的方法。根据语音信号和干扰噪声在离散余弦变换域DCT（discrete cosine transform）稀疏性的不同，提出一种基于改进压缩感知的说话人识别抗噪算法。在用正交匹配追踪OMP（orthogonal matching pursuit）算法重构语音信号时设定相关度阈值和语音恢复阈值，不仅有效恢复了语音信号，而且实现了语音增强。然后通过Gammatone滤波器组，对恢复语音信号进行处理，提取特征参数GFCC。仿真实验在高斯混合模型识别系统中进行，实验结果表明，将这种方法应用于说话人识别抗噪系统，系统的识别率及鲁棒性都有明显提高。%Compressive sensing (CS)is a method based on signal sparseness,and can efficiently extract useful information from signals. In this paper we present a speaker recognition anti-noise algorithm,which is based on improved compressive sensing,according to the different sparseness between speech signal and interfering noises in discrete cosine transform (DCT)area.We set correlation threshold and speech recovery threshold when reconstructing speech signals with orthogonal matching pursuit algorithm,this can not only restore speech signal effectively,but also realises the speech enhancement.Then through Gammatone filter bank we process the restored speech signal and extract feature parameter GFCC.Simulation experiment is conducted in Gaussian mixture model recognition system,experimental result shows that this algorithm obviously improves the recognition rate and robustness when being applied to speaker recognition and anti-noise system.