使用阈值提升基频特征对话者确认的补偿性能  

Threshold Used to Improve Compensation Performance of Pitch on Speaker Verification System

作　　者：花城[1] 李辉[1] 

机构地区：[1]中国科学技术大学电子科学与技术系,合肥230027

出　　处：《小型微型计算机系统》2014年第4期869-871,共3页Journal of Chinese Computer Systems

摘　　要：在传统的话者确认评分级融合中,基频辅助系统会劣化主系统下表现较好的测试评分.针对此问题,给出一种利用阈值改进基频特征补偿性能的方法.在使用MFCC特征参数的主系统外构建使用短时基频特征的话者确认辅助系统,主系统下测试评分绝对值低于阈值时需要进行评分融合,否则直接采用主系统评分.采用本文方法保护了主系统下的优质测试结果,实验结果表明,本文的方法有效提高了基频特征对说话人确认主系统的补偿性能.Previous fusion methods under speaker verification task degrade a part of better test scores of main system. In this paper, a new method using threshold decision is provided to improve the compensation performance of auxiliary system based on pitch. In addi- tion to a main system using MFCC, build an auxiliary system. A lot of test data is analyzed to prove the feasibility of using threshold. Linear weighted fusion between main system and auxiliary system is done when absolute test scores are below the threshold, or directly main system scores are used. Applying this method better scoring distribution of main system can be protected more effectively. Experi- mental results show that this method is effective to improve the compensation performance of pitch on speaker verification system.

关 键 词：阈值 基频 评分分布 补偿性能 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]