Separação automática de instrumentos de percussão brasileira a partir de mistura pré-gravada

Abstract

In this work, we present an investigation about blind source separation. Blind source separation, currently widely researched, is applied in many areas of knowledge, being here evaluated in the musical context, with Brazilian percussion instruments. The approached separation techniques are described: the NMF (nonnegative matrix factorization) along with its extension NMFD (non-negative matrix factor deconvolution), and the spectrogram median ltering which are essentially di erent, but combined in this thesis to separate percussive mixture signals. The term mixture identi es an audio track where two or more instruments are playing together. The two main contributions of this dissertation are: the investigation of blind source separation techniques applied to the signals' subbands and a recorded database of Brazilian percussion instruments. Other source separation systems are proposed; they combine the use of subbands, the NMF, and the median lter in various arrangements to enhance the mixture separations results. Futhermore, the e ect of the mixture quality on the system performance are evaluated. Mixtures with high reverberation times have been produced with the aid of a professional simulation software to test the proposed systems.