Overlapped Speech Detection for Improved Speaker Diarization on Tamil Dataset

Abstract

Speaker diarization is the task of partitioning a speech signal into homogeneous segments corresponding to speaker identities. We introduce a Tamil test dataset, considering that the existing literature on speaker diarization has experimented with English to a great extent; however, none on a Tamil dataset. An overlapped speech segment is a part of an audio clip where two or more speakers speak simultaneously. Overlapped speech regions degrade the performance of a speaker diarization system proportionally due to the complexity of identifying individual speakers. This study proposes an overlapped speech detection (OSD) model by discarding the non-speech segments and feeding speech segments into a Convolutional Recurrent Neural Network model as a binary classifier: single speaker speech and overlapped speech. The OSD model is integrated into a speaker diarizer, and the performance gain on the standard VoxConverse and our Tamil datasets in terms of Diarization Error Rate are 5.6% and 13.4%, respectively.