Automatic musical genre classification can be defined as the science of finding computer algorithms that a digitized sound clip as input and yield a musical genre as output. The goal of automated genre classification is, of course, that the musical genre should agree with the human classificasion. This demo illustrates an approach to the problem that first extract frequency-based sound features followed by a "linear regression" classifier. The basic features are the so-called mel-frequency cepstral coefficients (MFCCs), which are extracted on a time-scale of 30 msec. From these MFCC features, auto-regressive coefficients (ARs) are extracted along with the mean and gain to get a single (30 dimensional) feature vector on the time-scale of 1 second. These features have been used because they have performed well in a previous study (Meng, Ahrendt, Larsen (2005)). Linear regression (or single-layer linear NN) is subsequently used for classification. This classifier is rather simple; current research investigates more advanced methods of classification.
|Place of Publication||Kgs. Lyngby|
|Publication status||Published - 2005|