Automatic Multiscale-based Peak Detection on Short Time Energy and Spectral Centroid Feature Extraction for Conversational Speech Segmentation.

In this paper, we present a conversational speech segmentation system. We assume that the speech/non-speech has different energy in time and frequency domain. Therefore, the short time energy and spectral centroid are proposed as the feature extraction technique and the automatic multiscale algorithm as the signal peak detection. In the pre-processing phase, the Savitzky-Golay filter is performed to reduce the noise before feature extraction process. The short time feature serves to capture a short-stroke character and the spectral centroid feature is used to response the spectral characteristic. For observation, experimental and validation process, we use the six conversations of the SUSAS Database. The experimental result shows that the average accuracy of the proposed system is 98.26%.