基于韵律特征的说话人确认系统融合研究  

Fusion of Speaker Verification System based on Super-segment Prosodic Feature

作　　者：童强[1] 李辉[1] 方昕 

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 [2]科大讯飞信息科技股份有限公司,安徽合肥230088

出　　处：《通信技术》2013年第11期90-94,共5页Communications Technology

摘　　要：提出一种基于超音段韵律特征和GMM-UBM-MAP的文本无关的说话人确认系统,并与基于MFCC特征参数的说话人确认系统融合,研究提出新的两系统融合策略。在超音段中提取基于基频的韵律特征参数,建立辅助系统。融合时,以基准系统基于MFCC特征参数的说话人确认系统为主系统,基于韵律特征参数的系统为辅助系统,当主系统的得分与阈值接近时,将两系统得分融合再判断。通过NIST2006数据库的实验表明,融合系统相对原系统有16.39%的提升。A test-independent speaker verification method based on super-segment prosodic feature and GMM-UBM-MAP is proposed, and fused with another system based on MFCC, a fusion strategy for two systems is suggested. With the extraction rhythm feature parameter based on pitch, the auxiliary system is constructed. In this fusion, the main system is based on MFCC, while the auxiliary system based on super -segment prosodic feature. When the score of the main system is close to threshold, the scores of these two systems are fused, and then the judgement is done. Experiments show that the equal error rate （EER） of the auxiliary system can reach 17.77% , and the performance of final system is improved by 16.39% as compared with the main system.

关 键 词：超音段韵律特征 说话人确认 得分融合 

分 类 号：TN912.34[电子电信—通信与信息系统]