Feature contours fusion for determining segment boundaries in audio data

In this paper, an approach to audio segmentation based on fusion of feature trajectories is presented. The proposed scheme uses the variability of envelopes calculated from feature contours to determine change-points in audio stream. The contours are calculated by comparing adjacent frames utilizing distance or divergence functions. Such technique with the selected feature type can emphasize the change-points structure in the data. From calculated trajectory a Hilbert envelope is computed and the peaks and valleys are detected. As obtained results show, the positions of peaks and valleys are close to the actual position of segment boundaries. However, such situation leads to many miss and false alarm errors. Therefore, in our approach we have used a simple fusion based on averaged position with defined tolerance of several feature contours. We have performed an analysis of many features and functions to determine pairs with the high discriminatory power. In the result, for prepared audio stream we determined features and functions to obtain high accuracy of segmentation. The results show that utilizing our technique on audio with several audio classes can improve final detection accuracy.