Audio content analysis based on density of peaks in amplitude envelope

This paper presents an approach to audio parameterization using properties of the peaks detected in the amplitude envelope. The proposed solution based on observation that abrupt changes in the envelope of signal are connected with type of audio signal. For this purpose we used the density properties of peaks to calculate the feature vectors. The extraction process exploits an amplitude envelope estimation of signals obtained at the output of the filter bank. At the evaluation stage, three types of popular filter banks like mel-scale, linear-scale and gammatone were used and compared. As an example, we have performed popular speech/music classification task to determine the quality of proposed features. The influence of specified filter bank on the accuracy depends on the number of filters that covers different bands. The experimental results show that proposed descriptors have discrimination ability, very low computational complexity and can be exploited in audio classification tasks.