Pre-trained Deep Neural Network using Sparse Autoencoders and Scattering Wavelet Transform for Musical Genre Recognition

This paper tries to combine the approach of Deep Neural Networks (DNN) with the novel audio features extracted using the Scattering Wavelet Transform (SWT) for classifying musical genres. The SWT uses a sequence of Wavelet Transforms to compute the modulation spectrum coefficients of multiple orders which was already shown to be promising for this task. The DNN in this work uses layers pre-trained using Sparse Autoencoders (SAE). Data obtained from the Creative Commons website jamendo.com is used to boost the well-known GTZAN database, which is a standard benchmark for this task. The final classifier is tested using a 10-fold cross validation to achieve results similar to other state-of-the-art approaches.