Abstract:

In this report we present a new approach for instrument recognition in polyphonic multi-instrumental music. This work was done in the context of French ANR project SARAH ("StAndardisation du Remastering Audio Haute définition"). The main goal of this project is to develop audio source separation methods that would be efficient and applicable to a wide variety of music recordings. Thus, in the context of the project, the role of instrument recognition system is to locally identify the instruments of a music piece, in order to simplify the choice of a priori knowledge (e.g., expressed by probabilistic models of sources) used for separation. Our original approach for instrument recognition in polyphonic multi-instrumental music is based on the following steps:

1. signal decomposition into spectral components using Non-negative Matrix Factorization (NMF),
2. maximum a posteriori (MAP) estimation of "instrumental" components of the mix (adaptive Wiener filtering),
3. extraction of features from estimated (separated) components. However, in order to increase robustness of this feature extraction step, we introduce an original feature weighing mechanism trying to exclude unlikely features, i.e., those extracted from "badly separated" elementary components.
4. Support Vector Machine (SVM) classification.

Several configurations of the proposed instrument recognition system are evaluated on a jazz music database.

Back to Alexey Ozerov's Home Page