摘要
在声纹识别系统中,对声纹参数的提取很重要。传统的MFCC参数忽略了语音信号的动态特性,因此提出了一种基于小波包变换的声纹参数提取方法。为了更突出说话人的声纹特征,克服说话内容不同对提取声纹参数的影响,在分帧阶段采用帧长为2560点,增长有效语音段。再结合基于矢量量化(VQ)系统进行说话人识别实验,并通过比较常用的db3、db4、db6、coif3小波函数选取最优基。实验证明,相对于常用的256点帧长,帧长为2560点的识别率较高且提高了运算速率。coif3小波函数为声纹参数提取的最优基。新的WPT参数的识别率优于传统的MFCC参数。
In speaker recognition system, the voice parameters extraction is very important. The traditional MFCC parameter ignores the dynamic characteristics of speech signal, so a method is presented for extracting voice parameters based on wavelet packet transform. Text independent voice recognition system is to voice a more prominent feature of the speaker and overcomes the different speech content effects on the voiceprint parameters extraction. The frame length is adopted to increase effective voice for 2560 points in framing stage. And vector quantization （VQ） is combined with the speaker recognition experiment system, through the comparison of db3, db4, db6, coil3 wavelet function to choose the best basis. Experimental results show that frame length within 2560 points is higher and improves com- puting speed in comparison with common 256 point of the frame length. The optimal base coif3 wavelet function is taken as voiceprint parameter extraction. The MFCC parameter i- dentification of the WPT parameters of the new rate is better than tradition one.
出处
《沈阳理工大学学报》
CAS
2015年第6期77-82,共6页
Journal of Shenyang Ligong University
关键词
声纹参数
小波包变换
能量
矢量量化
语音信号
voiceprint parameter
wavelet transform
energy
vector quantization
speech signal
