一种三层判决的说话人索引算法

Speaker Index Algorithm of Three-layer Criterion

为提高说话人索引准确率,提出一种三层判决的说话人索引算法。第1层使用惩罚距离公式对说话人改变进行检测,第2层采用说话人模型自举法进行初次说话人辨认,第3层采用GMM说话人超级矢量进行判决,解决说话人模型自举法中产生的数据不匹配问题。实验结果表明,采用惩罚距离公式,与贝叶斯信息判决方法相比不需调整参数,与DISTBIC方法相比F1值提高2%,使用GMM说话人超级矢量,在说话人索引准确率和数量准确率方面分别提高8.95%、18.25%。

To improve the precision of speaker index,a speaker indexing algorithm of three-layer criterion is proposed.In the first layer,penalty distance is proposed to judge whether speaker changes.In the second layer,speaker model bootstrapping is used to identify speaker first time.In the third layer,GMM Speaker Supervector（GMMSS） is used to identify speaker further in order to settle the problem of data mismatch in speaker model bootstrapping.Experimental results show that,it is no need to tune penalty factor compared to BIC and F1 can improve 2% compared to DISTBIC;speaker indexing accuracy can improve 8.95% and the accuracy on the number of speaker can improve 18.25% by using GMMSS in speaker identification.