Self-optimized spectral correlation method for background music identification

This paper proposes a new method of detecting a known reference signal in an input signal highly corrupted by other sounds. One major application of the method is the identification of broadcast background music corrupted by speech. In this method, the reference signal is first decomposed into a number of small time-frequency components, and the maximum similarity between each component and the input is calculated. The similarities for all the components are then integrated by a voting method. Finally, the result is used to determine whether or not the reference exists in the input; and if it exists, to determine its position. Experiments on the identification of background music and the classification of similar TV commercials have shown that this method can identify 100% of target signals with an SNR of −10dB.