Hybrid approach for unsupervised Audio Speaker Segmentation

This paper deals with a new technique, DIS_T²_BIC, for audio speaker segmentation when no prior knowledge of speakers is assumed. This technique is based on a hybrid concept which is organized in two steps: the detection of the most probable speaker turns and the validation of turns already detected. For the detection our new technique uses a new distance measure algorithm based on the Hotelling´s T²-Statistic criterion. The validation is obtained by applying the Bayesian Information Criterion (BIC) segmentation algorithm to the detected speaker turns. For measuring the performance we compare the segmentation results of the proposed method versus recent hybrid techniques. Results show that DIS_T²_BIC method has the advantage of high accuracy speaker change detection with a low computation cost.