Unsupervised help-trained LS-SVR-based segmentation in speaker diarization system

摘要

In this paper, we propose a new segmentation method for diarization applications. In the proposed method, segmentation is performed using a discriminatively trained support vector regression, while a generative classifier helps it to estimate the probable change points. Since, there is no pre-labeled training samples in segmentation task, the proposed model-based segmentation method tries to suggest a proper solution to bridge this gap. It is assumed that initial applied samples are labeled with the first speaker in an unsupervised manner, while the subsequent training samples are chosen by applying the help-training approach. These samples are estimated to be conducive when both regression and classifier blocks, label positiveegative samples to be advantageous. These samples would be purified in next steps and speakers' models would be updated iteratively. In addition, a new procedure is introduced to estimate deleted and inserted change points that is executed when segmentation is completed. In comparison to similar approaches, experiments have shown performance improvement about 29% in diarization error rate.