Musical Cross Synthesis using Matrix Factorisation

Abstract

The focus of this work is to explore a new method for the creative analysis and manipulation of musical audio content. Given a target song and a source song, the goal is reconstruct the harmonic and rhythmic structure of the target with the timbral components from the source, in such a way that so that both the target and the source material are recognizable by the listener. We refer to this operation as musical cross-synthesis. For this purpose, we propose the use of a Matrix Factorisation method, more specifically, Shift-Invariant Probabilistic Latent Component Analysis (PLCA). The input to the PLCA algorithm are beat synchronous CQT basis functions of the source whose temporal activations are used to approximate the CQT of the target. Using the shift invariant property of the PLCA allows each basis function to be subjected to a range of possible pitch shifts which increases the flexibility of the source to represent the target. To create the resulting musical cross-synthesis the beat synchronous, pitch-shifted CQT basis functions are inverted and concatenated in time.