Including parametric models in spectrogram decomposition.

Abstract

Regarding the decomposition of an audio stream in elementary sound objects, the non-negative matrix factorization (NMF) is now one of the prominent tools that can be employed. For instance, in the context of an automatic music transcription application, the interest of the technique relies on its ability to use redundancies over a whole piece to hopefully be able to identify separated sound objects as musical notes. In this ideal case, the spectrogram factorizes into a matrix of inferred spectral atoms and a matrix of temporal activations associated to these atoms. However, the unconstrained version of the NMF does not guarantee that the resulting atoms are actually musical notes: the spectrum of a single note can be split into several atoms or, conversely, a single atom could represent more than one note. The work that will be presented here deals with parametric models designed to describe the spectral modifications that occur along the progress of a single audio event (e.g., a musical note) in order to obtain a more meaningful decomposition from a musical point of view.