Improving score-informed source separation for classical music through note refinement

Signal decomposition methods such as Non-negative Matrix Factorization (NMF) demonstrated to be a suitable approach for music signal processing applications, including sound source separation. To better control this decomposition, NMF has been extended using prior knowledge and parametric models. In fact, using score information considerably improved separation results. Nevertheless, one of the main problems of using score information is the misalignment between the score and the actual performance. A potential solution to this problem is the use of audio to score alignment systems. However, most of them rely on a tolerance window that clearly affects the separation results. To overcome this problem, we propose a novel method to refine the aligned score at note level by detecting both, onset and offset for each note present in the score. Note refinement is achieved by detecting shapes and contours in the estimated instrument-wise time activation (gains) matrix. Decomposition is performed in a supervised way, using training instrument models and coarsely-aligned score information. The detected contours define time-frequency note boundaries, and they increase the sparsity. Finally, we have evaluated our method for informed source separation using a dataset of Bach chorales obtaining satisfactory results, especially in terms of SIR.