Single Channel Beatbox Music Separation Using Non-negative Matrix Factorisation

Abstract

In this work, Non-negative Matrix Factorization (NMF) based single-channel beatbox music separation with spectral masking is proposed. Beatbox is an instrument like music entirely produced by the human voice. We have a mixed-signal of two musical beat sources. The algorithm uses data training of these two different beat sources with the NMF technique followed by spectral masking to separate an individual beat source out of the mixed-signal. During the training stage, we have used NMF in the magnitude spectrum domain to train the set of basis vectors for each beat source. After the observation of the mixed-signal, its magnitude spectra are decomposed into a linear combination of the trained basis for both beat sources. With the help of these decomposed spectrograms, masking is done which determines the role of each beat source in the mixed-signal and that is how we can separate and play a particular beat or a mixture of beats as per our musical sense and hence more versatile music can be created.