A non-parametric estimation approach for spectrum entropy computation in music-fingerprinting system

In this paper, we propose an approach for music-fingerprint, which can be obtained using Shannon´s entropy in the frequency domain for 24 critical bands of bark scale. We employ the non-parametric Parzen Window estimation to obtain the underlying Probability Distribution Function (PDF) of music clips in the beginning. Then, based on the PDF distribution estimated, we compute Shannon´s entropy in the frequency domain for 24 critical bands using a frame size of 20 ms with 50% overlapping. Once binary coded according to the certain algorithm, we can get the music-fingerprint of music clips. The experimental results show the robustness of audio fingerprint derived from our proposed approach, compared to that derived from the parametric method, under the classical degraded circumstances (i.e. additive noisy, low-pass filtering and lossy compression situation, etc.).