Music Signal Processing with Emphasis on Genre Classification

Abstract

Distribution estimation of music signals is necessary both for analysis and synthesis tasks. Genre classification is also one of the most fundamental problems in music signal processing. The present work is an effort to understand the probability distribution of music signals with an aim to classify music genres. For this four well known speech distributions viz. Gaussian, Generalized Gamma, Laplacian and Cauchy have been tested as hypotheses. The distribution estimation has been carried out in time domain, DCT domain and wavelet domain. It was observed that skewed Laplacian distribution describes the music samples most accurately with the peakedness of the distribution being correlated with the genre of music. Although Cauchy distribution along with Laplacian has been a good fit for most of the data, it is analytically shown in this work that Laplacian distribution is a better choice for modeling music signals. Genre classification between Metal and Rock genres was achieved with 78% accuracy using wavelet denoising.