说话人识别系统中特征提取的优化方法  

Feature extraction optimization method in speaker recognition system

作　　者：李铮 欧阳贝贝 赵淼 李琳 洪青阳[2] 童峰[3] LI Zheng;OUYANG Beibei;ZHAO Miao;LI Lin;HONG Qingyang;TONG Feng(School of Electronic Science and Engineering,Xiamen University,Xiamen 361005,China;School of Informatics,Xiamen University,Xiamen 361005,China;College of Ocean and Earth Sciences,Xiamen University,Xiamen 361102,China)

机构地区：[1]厦门大学电子科学与技术学院,福建厦门361005 [2]厦门大学信息学院,福建厦门361005 [3]厦门大学海洋与地球学院,福建厦门361102

出　　处：《厦门大学学报（自然科学版）》2020年第6期995-1003,共9页Journal of Xiamen University：Natural Science

基　　金：国家自然科学基金(61876160)。

摘　　要：声学特征提取是语音信号处理,如语音识别、语音唤醒、说话人识别等的核心技术之一.围绕说话人识别任务详细介绍了其主流声学特征,以及采用这些声学特征后所产生的问题和改进方法.同时,基于说话人区分向量(x-vector)说话人识别架构提出了一种双声学特征整合的方法,以改善识别性能,并在公开的VoxCeleb1数据集上对比几种常用声学特征在x-vector框架下的说话人识别结果,进一步验证了所提出的双特征整合方法的有效性.In the field of speech signal processing,including speech recognition,speech wake-up,and speaker recognition among others,the extraction of acoustic features is regarded as one of the core technologies.In this paper,mainstream acoustic characteristics of speaker recognition as well as draw-backs and improvement strategies are introduced in detail.Based on the x-vector speaker recognition architecture,this paper proposes the double feature integration architecture to improve the system performance.On the public VoxCeleb1 dataset,we compare speaker recognition results of several commonly acoustic features in the x-vector framework,and validate the effectiveness of the presented double features integration method.

关 键 词：说话人识别 声学特征 特征优化 双特征 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]