Exploring different strategies for music genre classification

Presenter: Antonio José Homsi Goulart

The possibility of streaming and the ease to download and store music in computers and portable devices make AMGC (automatic music genre classification) systems a must. Systems based on metadata analysis might be unprecise, and classifications are by artist or by album, rather than by each tune. Other systems, such as those that addopt MFCCs, explores rhytmic and timbric content of the songs, being LDA, SVM and GMM the most used classifiers. The main goal of this work is to investigate if wavelet based entropy, fractal dimension and lacunarity are good parameters to represent music signals and provide a good accuracy in an AMGC system based on these concepts, rather than musical information. The classifiers adopted were SVMs and GMMs. Two databases were created for the tests; one based on 3 genres (blues/classical/lounge) and the other on 4 brazilian genres(axé/bossa-nova/forró/samba).

When: September 27th, 2011

Where: Room 259-A, at IME/USP