改进的基于决策树的说话人在线聚类

Improved online speaker clustering based on decision tree

针对采用传统的在线聚类方法时后续判决错误率较高的缺点,提出了一种改进的基于决策树的在线说话人聚类算法。通过构建一个决策树,增加判决分支,对语音段进行判决聚类,从而有效降低前期错误判决对后续聚类的影响。为了进一步提高算法效率,缩短运算时间,还给出了一种决策树剪枝方法,减少了不合理的判决分支。通过对广播新闻语料进行的说话人聚类实验表明,相比传统的层次聚类算法,新算法的平均类纯度和说话人纯度分别提高了0.9%和1.1%,计算时间减少了57%。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 a key component in many speech processing applications.To solve the problem of error propagating in the posterior clustering caused by the traditional online clustering,an improved online speaker clustering algorithm based on a decision tree is proposed.Unlike typical online clustering approaches,the proposed method constructs a decision tree to increase branches and to distinguish an audio segment clustering to reduce effectively the effect of error distinguishing on the posterior clustering.To shorten the operation time,a pruning strategy for candidate-elimination is also presented.Experiments indicate that the algorithm achieves good performance on both precision and speed.By using this method,the average speaker purity and the average cluster purity have improved by 0.9% and 1.1% respectively,and the time consuming is reduced by 57%.Experiments also show that this method is effective for improving the performance of the unsupervised adaptation as compared with the true speaker-condition.