Towards Audio to Score Alignment in the Symbolic Domain

This paper presents a matrix factorization based feature for audio to score alignment. We show that in combination with dynamic time warping it can compete with chroma vectors, which are the probably most frequently used approach within the last years. A great benefit of the factorizationbased feature is its sparseness, which can be used in order to transform it into a symbolic representation. We will show that music to score alignments using the symbolic version of the feature is less accurate but on the other hand reduces the memory required for feature representation and during the alignment process to a fraction of the original amount. This is of special value when dealing with very long pieces of music where the limits of default DTW are reached.