论文部分内容阅读
说话人聚类是语音识别以及音频检索等众多语音应用的重要组成部分。提出一种改进的基于层次聚类的说话人聚类算法,对层次聚类法做出了进一步改进:(1)改进误差平方和准则以提高聚类速度;(2)引入假设检验方法确定类别数目;(3)提出一种稳健的在线聚类方法以解决对新到来的语音段进行聚类的问题。在聚类实验中,算法的平均类纯度和说话人纯度分别为96.7%和96.6%。实验结果还表明,相比手工标注说话人信息,将该算法的聚类结果应用于说话人自适应可降低系统的误识率。
Speaker clustering is an important part of many speech applications such as speech recognition and audio retrieval. This paper proposes an improved clustering algorithm based on hierarchical clustering to further improve the hierarchical clustering method: (1) to improve the squared error square rules to improve the clustering speed; (2) to introduce the hypothesis testing method to determine the category (3) A robust on-line clustering method is proposed to solve the problem of clustering newly arrived segments of speech. In the clustering experiments, the average class purity and speaker purity of the algorithm are 96.7% and 96.6% respectively. The experimental results also show that compared with the manual annotation of the speaker information, applying the clustering result of the algorithm to the speaker adaptation can reduce the error rate of the system.