用说话人相似度i-vector的非负值矩阵分解说话人聚类

摘要

Based on Bayesian or full Bayesian criterion, the speaker clustering or recognition method is mainly used to repeat the similarity measure of the whole utterance segment, and then combine the similar utterance segment to realize speaker clustering.In this method, if the number of utterance segment is increased, the combined computation time is longer and the system real-time property is worse.Moreover, the speaker model is established by GMM.The reliability of GMM is reduced when the speech time is short, which affects the accuracy of speaker clustering.Aiming at the above problems, this paper proposes a high-accuracy fast speaker clustering method based on non-negative matrix factorization and i-vector of speaker similarity.%基于贝叶斯或者全贝叶斯准则的说话人自动聚类或者识别方法,主要采取重复换算全发话语音段的相似量度,再组合相似性较大的语音片段实现说话人的聚类.这种方法中如果发话语音片段数越多,组合计算时间就越长,系统实时性变差,而且各说话人模型用GMM方法建立,发话语音时间短暂时GMM的信赖性降低,最终影响说话人聚类精度.针对上述问题,提出引用i-vector说话人相似度的非负值矩阵分解的高精度快速说话人聚类方法.