Speech Overlap Detection using Spectral Features and its Application in Speech Indexing

Simultaneous occurrence of speech from one or more than one speaker is considered as speech overlap. Speech overlap is due to distortion in transmission channels or vicinity of voice sources. Speech overlap has destructive effects in performance of speech recognition systems. Speech overlap detection is one of the main areas in speech and speaker indexing. In speaker indexing, speech signal is partitioned into segments where each segment is uttered by only one speaker. So, parts of speech that include two or more speakers simultaneously should be determined before any following processes. Speaker overlap detection is also useful in some other speech processing applications including speech and speaker recognition. In this paper we propose a simple and efficient method for speaker overlap detection. This method uses spectral periodicity of voiced frames to decide whether a speech segment includes speech overlap or not. In this method, voiced frames which do not have a periodic Fourier spectrum are labeled as speech overlaps. Experimental results show that the performance of the proposed method is relatively good. The main advantage of this method is its speed in labeling segments containing speech overlaps