Automatically Adapting the Structure of Audio Similarity Spaces

Today, among the best-performing audio-based music similarity measures are algorithms based on Mel Frequency Cepstrum Coefficients (MFCCs). In these algorithms, each music track is modelled as a Gaussian Mixture Model (GMM) of MFCCs. The similarity between two tracks is computed by comparing their GMMs. One drawback of this approach is that the distance space obtained this way has some undesirable properties. In this paper, a number of approaches to correct these undesirable properties are investigated. They use knowledge about the properties of music by using other music tracks as a reference. These reference tracks can either be the music collection itself, or they may be an external set of reference tracks. Our results show that the proposed techniques clearly improve the quality of this audio similarity measure. Furthermore, preliminary experiments indicate that the techniques also help to improve other similarity measures. They may even be useful in completely different domains, most notably text information retrieval.