Features of average spectral envelope for audio regions determination

A technique to determine change-points in audio stream is presented in this paper. Proposed approach exploits the properties of averaged spectral envelope. As was shown, the divergence of peaks positions and its values are connected with the type of source signal. Therefore, measured statistical indicators for each frame can be used to detect the change-points in the signal. The method considered in this paper has been verified for radio broadcast signal using manually labeled regions as reference. Performed experiments show that feature trajectories can indicate positions where the signal changes its classes (music, speech and mixtures). Selection of feature is crucial for final accuracy, thus several experiments for different extraction conditions have been performed. In most cases the positions of peaks are the most discriminative with moderate level of errors. Using discriminator procedure on obtained trajectory representation, one could extract the region boundaries of the audio waveform.