Speaker segmentation for intelligent responsive space

Abstract

Information drawn from conversational speech can be useful for enabling intelligent interactions between humans and computers. Speaker information can be obtained from speech signals by performing Speaker Segmentation. In this paper, a method for Speaker Segmentation is presented to address the challenge of identifying speakers even when utterances are very short (0.5sec). This method, involving the selective use of feature vectors, experimentally reduced the relative error rates by 27-42% for groups of 2 to 16 speakers as compared to the conventional approach for Speaker Segmentation. Thus, this new approach offers a way to significantly improve speech-data classification and retrieval systems.