Musical genre classification based on a highly-resolved cepstral modulation spectrum

We propose new features for musical genre classification which are based on the modulation spectrum of cepstral coefficients, and investigate the impact of the modulation frequency resolution on the classification accuracy. We compare the performance of the novel feature set which is derived from a high-resolution modulation spectrum to that of two feature sets which are either based on a coarsely resolved modulation spectrum or roughly summarize the modulation energy in a few bands. From the results of a 5-class musical genre classification experiment it can be concluded that a high modulation frequency resolution is crucial for representing the harmonic modulation structure of Electronic music in particular. The proposed features outperform the two competing methods with an overall detection rate of 81%. After computing the cepstral modulation spectrum with efficient FFT operations, the computational complexity for feature extraction is fairly low as only 22 low-level features need to be computed.