Blind Enhancement of the Rhythmic and Harmonic Sections by NMF: Does it help?

Non-Negative Matrix Factorization is well known to lead to considerable successes in the blind separation of drums and melodic parts of music recordings. Such splitting may well serve as enhancement when it comes to typical Music Information Retrieval tasks as automatic key labelling or tempo detection. In this respect we introduce the combination of an NMF based blind music separation into several isolated audio tracks in combination with Support Vector classification to assign each obtained track to be either rhythmic or melodic. Thereby optimal parametrization and performances are discussed. Next, stereophonic information is further used to eliminate the key melody and bass usually panned in the centre for tempo detection or e. g. for chord labelling. We then analyse the potential for the named tasks by a number of experiments carried out on the MTV Europe Most Wanted of the 1980 ies and 90 ies in MP3 format.