基于基音周期与清浊音信息的梅尔倒谱参数  

Mel-Cepstrum Integrated with Pitch and Information of Voiced/Unvoiced

作　　者：郭武[1] 王仁华[1] 戴礼荣[1] 

机构地区：[1]中国科技大学电子工程与信息科学系,合肥230027

出　　处：《数据采集与处理》2007年第2期229-233,共5页Journal of Data Acquisition and Processing

基　　金：中国科技大学青年教师基金;国家863(2006AA010104)资助项目

摘　　要：提出一种在浊音部分不固定帧长的梅尔倒谱参数(Mel-cepstrum)提取的方法。针对浊音和清音所包含信息量不同,对浊音进行双倍的加权,从而将基音与清浊音信息融合进梅尔倒谱参数。将这种动态的梅尔倒谱参数应用在说话人确认中,在混合高斯模型(Gaussian mixture models,GMM)的情况下,取得了比常用的梅尔刻度式倒频谱参数(Mel-frequency cepstral coefficient,MFCC)更高的识别率,在NIST 2002年测试数据库中,512个混合高斯下能够将等错误率(EER)由9.4%降低到8.3%,2 048个混合高斯下能够将等错误率由7.8%降低到6.9%。An algorithm that extracts the Mel-cepstrum using variable frame length during voiced speech is proposed. Furthermore, the voiced part Mel-cepstrum is copied twice because more information is held in voiced speech than in unvoiced speech. The information of pitch and voiced/unvoiced is fused into the Mel-cepstrum through the above two methods. When the Gaussian mixture models （GMM） is adopted in text-independent speaker verification, the system based on the dynamic Mel-cepstrum（DMCEP） has better performance than the system based on standard Mel-frequency cepstral coefficient （MFCC）. Speaker verification experiments are carried on the 2002 NIST single speaker verification evaluation corpus. Compared with standard MFCC, the equal error rate （EER） is reduced to 8. 3% from 9.4% and to 6.9% from 7.8% in 512 GMM and 2 048 GMM with DMCEP.

关 键 词：说话人确认 梅尔倒谱参数 基音频率 清浊音信息 

分 类 号：TN912.34[电子电信—通信与信息系统]