Independent Component Analysis for Music Similarity Computation

In the recent years, a number of publications have appeared that deal with automatically calculating the similarity of music tracks. Most of them are based on features that are not intuitively understandable to humans, as they do not have a musically meaningful counterpart, but are merely measures of basic physical properties of the audio signal. Furthermore, most of these algorithms do not take into account the temporal development of the audio signal, which certainly is an important aspect of music. All of them consider the musical signal as a whole, not trying to reconstruct the listening process of dividing the signal into a number of sources. In this work, we present a novel approach to fill this gap by combining a number of existing ideas. At the heart of our approach, Independent Component Analysis (ICA) decomposes an audio signal into individual parts that appear maximally independent from each other. We present one basic algorithm to use these components for similarity computations, and evaluate a number of modifications to it with respect to genre classification accuracy. Our results indicate that this approach is at least of similar quality as many existing feature extraction routines.