Multi-Instrumental Deep Learning for Automatic Genre Recognition

The experiments described in this paper utilize songs in the MIDI format to train Deep Neural Networks (DNNs) for the Automatic Genre Recognition (AGR) problem. The MIDI songs were decomposed into separate instrument groups and converted to audio. Restricted Boltzmann Machines (RBMs) were trained with the individual groups of instruments as a method of pre-training of the final DNN models. The Scattering Wavelet Transform (SWT) was used for signal representation. The paper explains the basics of RBMs and the SWT, followed by a review of DNN pre-training methods that use separate instrument audio. Experiments show that this approach allows building better discriminating models than those that were trained using whole songs.