Evaluation of speech music transitions in Radio programs based on acoustic features

The final target of our project is to create an automatic program editor for radio. There are several algorithms for music playlist generation, but no reference has been found for mixed speech and music playlists. As for a first step we studied the elements of existing radio programs. A simple subjective opinion test has been constructed to evaluate the ability of normal listeners to discriminate the well edited and the randomly selected speech and consecutive music pairs. Significant difference has been found in the opinions between the well harmonising pairs and the pairs having dissimilar characteristic. We tried to predict the opinion values based on the basic acoustic features of the speech and the music signals. Some relations can be established based on statistical methods in between the acoustic features and the opinion values. We hope that by using content based features and data mining methods this prediction can be more accurate.