Variable-Size Gaussian Mixture Models for Music Similarity Measures

An algorithm to efficiently determine an appropriate number of components for a Gaussian mixture model is presented. For determining the optimal model complexity we do not use a classical iterative procedure, but use the strong correlation between a simple clustering method (BSAS [13]) and an MDL-based method [6]. This approach is computationally efficient and prevents the model from representing statistically irrelevant data. The performance of these variable size mixture models is evaluated with respect to hub occurrences, genre classification and computational complexity. Our variable size modelling approach marginally reduces the number of hubs, yields 3-4% better genre classification precision and is approximately 40% less computationally expensive.