Unsupervised Speaker Clustering Using SVM Training Missclassification Rate for Short-Duration Speech Signals

This paper proposes an unsupervised speaker clustering system for duration of speech signals below 4 seconds. For determining whether two collected speech sections uttered from the same speaker or not, our previous SVM training miss-classification rate (STMR) is adopted to evaluate the data separability between two different speakers. This paper also proposes a hierarchical extract and merge (HEM) clustering method to reduce agglomeration time and enhance the clustering purity. Experiment results show the average speaker purity (ASP) and average cluster purity (ACP) are both better than the CE manner with the GMM training miss-classification rates (GTMR) for 2 to 4 seconds short speech sections.