计算机工程 ›› 2019, Vol. 45 ›› Issue (12): 263-266,273.doi: 10.19678/j.issn.1000-3428.0053482

• 多媒体技术及应用 • 上一篇    下一篇

基于AdaBoost算法的回放语音检测研究

贾甜博, 蒋晔   

  1. 南京财经大学 信息工程学院, 南京 210023
  • 收稿日期:2018-12-25 修回日期:2019-01-26 发布日期:2019-01-28
  • 作者简介:贾甜博(1992-),男,硕士研究生,主研方向为语音识别;蒋晔(通信作者),副教授、博士。
  • 基金项目:
    江苏省自然科学基金青年基金(BK20150987)。

Research on Playback Speech Detection Based on AdaBoost Algorithm

JIA Tianbo, JIANG Ye   

  1. College of Information Engineering, Nanjing University of Finance and Economics, Nanjing 210023, China
  • Received:2018-12-25 Revised:2019-01-26 Published:2019-01-28

摘要: 针对语音判别系统中单个分类器分类能力有限的问题,提出一种基于AdaBoost算法的回放语音检测方法。以常量Q倒谱系数和均值超矢量分别作为特征参数和AdaBoost算法的输入,将多个分类器的检测结果相结合并进行加权投票,从而降低系统的等错误率(EER)。研究关系因子、均值超矢量维数以及弱分类器数量对检测结果的影响,以设置系统的最优参数。实验结果表明,该检测方法在开发集和评估集上的EER值分别为4.17%和16.81%,相比GMM-ML方法分别降低了65%和44%。

关键词: 自动说话人确认, 回放语音检测, AdaBoost算法, 均值超矢量, 加权投票

Abstract: To address the limited classification ability of a single classifier in speech recognition system,this paper proposes a playback speech detection method based on AdaBoost.In this method,the constant Q cepstrum coefficient is taken as the characteristic parameter,and the mean supervector is taken as the inputs of characteristic parameter and AdaBoost algorithm respectively.Then,this paper combines the detection results of multiple classifiers for weighted voting,thus reducing the Equal Error Rate(EER) of the system.Further,this paper studies the effects of relationship factor,mean supervector dimension and the number of weak classifiers on the detection results,so as to set the optimal parameters of the system.Experimental results show that the EER values of the proposed method in development set and evaluation set are 4.17% and 16.81% respectively,which are 65% and 44% lower than those of the GMM-ML method.

Key words: Automatic Speaker Verification(ASV), playback speech detection, AdaBoost algorithm, mean supervector, weighted voting

中图分类号: