Source-separated audio input for accelerating convolutional neural networks

Convolutional neural networks have recently been found useful in music information retrieval tasks such as genre classification, music recommendation, and instrument detection. In these deep neural network applications, music data is often converted into a spectrogram and treated as a single-channel image input. In this work we enrich the semantic musical understanding of the network by using source separation to separate the single channel into multiple, meaningful channels. As a result, we improve accuracy and accelerate training. We evaluate this method in a genre classification task with a variety of CNN architectures. Compared to single-channel inputs, we achieve an average of a 2.7% increase in accuracy across all trained models using a 3-channel source-separated model and a 2.2% increase using a 4-channel model. This is due to quicker convergence of the loss function.