基于样本熵的语音/音乐识别  

Speech/music discriminator based on sample entropy

作　　者：杨松[1] 于凤芹[1] 

机构地区：[1]江南大学物联网工程学院,江苏无锡214122

出　　处：《计算机工程与应用》2012年第23期125-127,154,共4页Computer Engineering and Applications

基　　金：国家自然科学基金(No.61075008)

摘　　要：传统的MFCC及短时能量特征只反映了信号序列的静态特征,目前基于这些特征的语音/音乐识别率为79%～86%。样本熵可以反映信号序列中的新信息量的大小以及新信息量的变化程度。以样本熵作为特征对语音/音乐进行分类识别,提取混合信号的样本熵,计算每段信号样本熵的均值和方差,采用k均值聚类进行识别。仿真实验结果表明,基于样本熵的语音/音乐识别的识别率可提高到88.073%。Mel frequency cepstral coefficients and short time energy only reflect the static characteristics in signal sequence and the recognition rate of speech/music discrimination is 79%～86%.Sample entropy reflects the size and variational extent of new information in signal sequence.This paper conducts speech/music discrimination using sample entropy.The mean and variance of the sample entropy are calculated after extracting the sample entropy of mixed signals,then each signal is recognized by k-means cluster.Simulation experimental results show that the recognition rate of speech/music discrimination reaches 88.073% when using sample entropy.

关 键 词：语音/音乐识别 样本熵 K均值聚类 

分 类 号：TN912[电子电信—通信与信息系统]