Improving Accuracy of Polyphinic Music-to -Score Alignment.

This paper presents a new method to refine music-to-score alignments. The proposed system works offline in two passes, where in the first step a state-of-the art alignment based on chroma vectors and dynamic time warping is performed. In the second step a non-negative matrix factorization is calculated within a small search window around each predicted note onset, using pretrained tone models of only those pitches which are expected to be played within that window. Note onsets are then reset according to the pitch activation patterns yielded by the matrix factorization. In doing so, we are able to resolve individual notes within a chord. We show that this method is feasible of increasing the accuracy of aligned note’s onsets which are already aligned relatively near to the real note attack. However it is so far not suitable for the detection and correction of outliers which are displaced by a large timespan. We also compared our system to a reference method showing that it outperforms bandpass filtering based onset detection in the refinement step.