基于MFCC与基频特征贡献度识别说话人性别  

Speaker gender recognition based on combining the contribution of MFCC and pitch features

作　　者：庞程[1] 李晓飞[1] 刘宏[1] 

机构地区：[1]北京大学深圳研究生院深圳物联网智能感知技术工程实验室,广东深圳518055

出　　处：《华中科技大学学报（自然科学版）》2013年第S1期108-111,120,共5页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：基金项目国家自然科学基金资助项目(60875050;60675025);国家高技术研究发展计划资助项目(2006AA04Z247);深圳市科技计划资助项目(JCYJ20120614152234873;CXC201104210010A;JCYJ20130331144716089;JCYJ2013033114 4631730)

摘　　要：提出了一种复杂场景下基于MFCC与基频特征贡献度的说话人性别识别方法.该方法有效融合了基于Mel频率倒谱系数的模板匹配方法和基音频率判别方法.实验语音数据库包括5 000个孤立词语音和1 260个带情感的语音.在安静环境下说话人的性别识别率可以达到98.88%,在信噪比为10dB的babble噪声下通过谱减法降噪后的识别率为90.2%.实验表明:说话人情绪对性别识别的影响较大,尤其是男声.A speaker gender identification method was presented based on the contribution of MFCC(Mel frequency cepstral coeffcients)and pitch under complex scenarios.This method effectively combines the Mel frequency cepstral coefficient template matching method and discrimination method based on pitch.The voicebox of this system includes 5 000isolated word speeches and 1 260speeches with emotion.The recognition rate of speaker gender in quiet environment can reach 98.88%,and the rate of the speaker gender after spectral subtraction is 90.2% under the babble noise environment when the SNR is 10dB.The experiment also indicates that emotion has a great effect on speaker gender recognition,especially male voice.

关 键 词：语音识别 说话人性别识别 信息融合 模板匹配 基音频率判别 

分 类 号：TN912.3[电子电信—通信与信息系统]