Discrete wavelet packet transform and ensembles of lazy and eager learners for music genre classification

This paper presents a process for determining the music genre of an item using a new set of descriptors. A discrete wavelet packet transform is applied to obtain the signal representation at two different resolutions: a frequency resolution and a time resolution tuned to encode music notes and their onset and offset. These features are tested on a number of data sets as descriptors for music genre classification. Lazy learning classifiers (k-nearest neighbor) and eager learners (neural networks and support vector machines) are applied in order to assess the classification power of the proposed features. Different feature selection techniques and ensemble methods are explored to maximize the accuracy of the classifiers and stabilize their behavior. Our evaluation shows that these frequency descriptors perform better than a standard approach based on Mel-Frequency Cepstral Coefficients and on the Short Time Fourier Transform in music genre classification. Moreover, our work confirms that a parameterization of the music rhythm based on the beat-histogram provides some meaningful information in the context of music classification by genre.Finally, our evaluation suggests that multi-class support vector machines with a linear kernel and round-robin binarization are the simplest and more effective process for music genre classification.