• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

J4 ›› 2012, Vol. 34 ›› Issue (6): 79-82.

• 论文 • 上一篇    下一篇

一种基于性别的说话人索引算法

杨继臣1,何俊2,李艳雄2   

  1. (1.仲恺农业工程学院计算机科学与工程学院,广东 广州 510225;
    2.华南理工大学电子与信息学院,广东 广州 510640)
  • 收稿日期:2011-03-26 修回日期:2011-06-28 出版日期:2012-06-25 发布日期:2012-06-25
  • 基金资助:

    广东省自然科学基金资助项目(10451064101004651);中央高校基本科研业务费专项资金资助项目(2011ZM0029)

A GenderBased Algorithm of Speaker Indexing

YANG Jichen1,HE Jun2,LI Yanxiong2   

  1. (1.School of Computer Science and Engineering,
    Zhongkai University of Agriculture and Technology,Guangzhou 510225;
    2.School of Electronics and Information Engineering,
    South China University of Technology,Guangzhou 510640,China)
  • Received:2011-03-26 Revised:2011-06-28 Online:2012-06-25 Published:2012-06-25

摘要:

为了提高说话人索引准确率,对说话人改变判决中常用的贝叶斯信息判决(BIC)进行改进和在说话人辨认中使用性别信息,提出了一种基于性别的说话人索引算法。首先使用惩罚距离公式对说话人改变进行检测,解决了在说话人改变判决中使用 BIC需要不断调节惩罚因子的问题;其次在说话人改变检测的基础上,采用性别模型判断每个说话人的性别;最后把男性和女性说话人分别对待,使用说话人模型自举法对说话人进行辨认。实验结果表明:在说话人改变检测中,采用惩罚距离公式,和BIC相比不需要调整参数,和DISTBIC相比,在F1方面提高了2%;在说话人辨认方面,利用性别信息,说话人索引准确率(SIA)提高了20.93%,说话人数量准确率(SNA)方面提高了3%。

关键词: 说话人索引, 性别信息, 说话人模型自举法

Abstract:

To improve the precision of speaker indexing, an algorithm of speaker indexing based on gender is proposed by modifying BIC which is often used in speaker change criterion and by using gender information. In the first step, penalty distance is proposed to judge whether a speaker changes, which settles the problem that it is needed  to tune the penalty factor of the Bayesian information criterion(BIC) repeatedly by the speaker change criterion. In the second step, a gender model is used to judge every speaker’s gender on the basis of speaker changes. In the third step, speaker model bootstrapping is used to identify a male speaker or a female speaker separately. The experimental results show that: it is unnecessary to tune the penalty factor compared to BIC and F1 is improved by 2% compared to DISTBIC by the  speaker change detection; the speaker indexing accuracy is improved by 20.93% and the accuracy on the number of speakers is improved by 3% by using the gender information in speaker identification.

Key words: speaker indexing;gender information;speaker model bootstrapping