Audio similarity measure based on Renyi's quadratic entropy

Considering noise interference often exists in audio processing, it is not robust enough to calculate audio similarity by using distance measure directly. In this paper, basing on Renyi's quadratic entropy, a novel scheme for audio similarity measure is proposed. In our work, we extract Mel Frequency Cepstral Coefficients (MFCCs) to represent each audio, and then calculate the similarity based on the entropy of audio samples by probability density function (pdf) of MFCCs which can be estimated by Parzen window. The experimental results show that: (a) our approach has better performance than the one based on Euclidean distance in the common SNR condition, (b) our approach can achieve 94.00% matching accuracy even when the signal to noise ratio (SNR) is 0db. In addition, our algorithm also can be applied in audio retrieval and musical cluster.