Cross-synthesis Based on spectrogram Factorization

Spectrogram factorization techniques decompose a sound into a set of characteristic spectral shapes and a set of corresponding temporal evolutions. This can be exploited for a cross-synthesis-like processing by combining the spectral shapes of one sound with the temporal evolutions of the other. A system is proposed that implements such a task in an unsupervised way by means of a comparison of the involved spectral shapes in terms of timbral similarity, and a phase reconstruction algorithm for the resynthesis. The system enables cross-synthesis at the level of intra-note resonances, transients or temporalities. Some illustrative sound generation examples will be presented and discussed.