Unsupervised Feature Pre-training of the Scattering Wavelet Transform for Musical Genre Recognition

This paper examines the utilization of Sparse Autoencoders (SAE) in the process of music genre recognition. We used Scattering Wavelet Transform (SWT) as an initial signal representation. 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 Autoencoders can be used for pre-training a deep neural network, treated as an features detector, or used for dimensionality reduction. In this paper, SAEs were used for pre-training deep neural network on the data obtained from jamendo.com website offering music on creative commons licence. The pre-training phase is performed in unsupervised manner. Next, the network is fine-tuned in supervised way with respect to the genre classes. We used GTZAN database for fine-tuning the network. The results are compared with those obtained with training neural network in a standard way (with random weights initialization).