Score-Informed Sparseness for Source Separation

Audio source separation is a useful preprocessing step for remixing or transcription of music. It can be shown, that the separation quality increases, if the separation algorithm gets additional side information, e.g. the score of the current mixture [5]. In many cases the score of a musical piece is not available and has to be extracted by a professional musician or an automatic music transcription algorithm. To avoid both necessities, we will propose a source separation algorithm, which utilizes only the temporal activity (TA) of each instrument in the mixture. Compared to the whole score, this TA can be evaluated with much less experience in music transcription. To improve separation quality, the TA controls the sparsity of a non-negative tensor factorization. We will show, that for certain mixtures, this TA is a sufficient information for source separation. If TA is not sufficient, it can be utilized as a preprocessing step for further blind source separation algorithms.