基于超音段韵律特征和GMM-UBM的文本无关的说话人识别  

Text-independent speaker recognition based on super-segment prosodic feature and GMM-UBM

作　　者：许东星[1] 戴蓓蒨[1] 刘青松[1] 许敏强[1] 

机构地区：[1]中国科学技术大学电子科学与技术系,安徽合肥230027

出　　处：《中国科学技术大学学报》2010年第2期157-162,共6页JUSTC

摘　　要：提出一种采用超音段韵律特征和GMM-UBM模型结构的文本无关的说话人识别方法,用多尺度小波分析方法从短时倒谱参数MFCC和基频F0随时间变化的韵律中分别提取可用于文本无关说话人识别的超音段韵律特征参数PMFCC和PF0,并组成联合参数PMFCCF0.在NIST068side-1side复杂背景电话手机语音数据库上的说话人确认实验则表明,采用一阶小波分析方法提取的超音段韵律参数PMFCC的识别性能与短时MFCC相当,采用超音段韵律特征PMFCCF0的系统确认性能比采用短时MFCC系统有较大的提高.在微软数据库进行不同信噪比测试语音的说话人辨认实验表明,PMFCCF0有比短时MFCC更好的噪声鲁棒性.A text-independent speaker recognition method was proposed based on the super-segment prosodic feature and GMM-UBM. With wavelet multiresolution analysis, the super-segment prosodic feature PF0 from Fo-t and PMFCC from MFCC-t were extracted, which were used for text-independent speaker recognition and could be combined as PMFCCF0. Experiments of speaker identification in different SNRs on Microsoft database indicate that PMFCCF0 is more robust than MFCC. Experiments on the 2006 NIST 8side-lside subset speaker recognition evaluation task show that PMFCC performs quite as well as MFCC in speaker recognition and the system verification performance based on PMFCCF0 exhibits better noise robustness compared with MFCC.

关 键 词：超音段韵律特征 GMM—UBM 文本无关 说话人识别 

分 类 号：TN912.34[电子电信—通信与信息系统]