基于改进Mel滤波器组的声纹特征提取SoC设计  

SoC Design of Voiceprint Features Extraction Based on Improved Mel Filter Banks

作　　者：陶佰睿[1,2] 郭琴[2] 苗凤娟[2] 李青龙[2] 张景林[1] 

机构地区：[1]齐齐哈尔大学计算中心,黑龙江齐齐哈尔161006 [2]齐齐哈尔大学通信与电子工程学院,黑龙江齐齐哈尔161006

出　　处：《微电子学》2015年第6期785-788,共4页Microelectronics

基　　金：国家自然科学基金资助项目(61204127);黑龙江省自然科学基金资助项目(F201332;F201438);黑龙江省普通高等学校新世纪优秀人才培养计划(1253-NECT025);黑龙江省教育厅科学技术研究项目(12541899;12531774)

摘　　要：针对说话人声纹身份认证一体化设计需求,利用曲线拟合方法改进Mel滤波器组,开展声纹特征参数提取的SoC设计研究。首先选取MFCC,IMFCC,MidMFCC的Mel频率刻度上的平均值作为拟合参数,得到拟合函数;然后利用插值方法改进Mel滤波器组中各滤波器的加权中心频率;通过Matlab进行算法设计和仿真分析,采用Quartus II和Modelsim实现代码编写、编译、时序仿真和综合;最后利用SOPC Builder技术将Verilog模块封装为IP核,并下载到DE2开发板中进行验证。研究结果表明,本系统提取出的多维特征参数具有更好的鲁棒性,有利于提高识别率。For voiceprint authentication integrated design requirements, the SoC design of Mel filter banks based voiceprint feature extraction which was improved by using curve fitting was presented. Firstly, the average value on Mel frequency scale of MFCC, IMFCC, MidMFCC was taken as a fitting parameter to get a fitting function. Then, the interpolation method was used to improve the weighted center frequency of each filter which was a part of Mel filter bank. The algorithm design and simulation were conducted by Matlab, and the coding, compilation, timing simulation and synthesis were implemented respectively through Quartus II and Modelsim. Finally, the verilog module was packaged to be an IP core by using SOPC Builder technology and downloaded to the DE2 development board for verifying. The results showed that multi-dimensional feature parameters extracted by the proposed circuit had better robustness, which was good for improving the recognition rate.

关 键 词：声纹身份认证 改进Mel滤波器组 声纹特征提取 片上系统 

分 类 号：TN492[电子电信—微电子学与固体电子学]