Large Scale Music Genre Classification with Deep Neural Networks

In this work, we focus on music classification, specifically music genre classification. We explore the performance of various neural network structures including deep long short-term memory recurrent neural networks (RNNs), deep convolutional LSTM RNNs, and deep convolutional neural networks (CNNs) utilizing both conventional convolutional layers as well as dilated convolutional layers with exhaustive experiments. We also compare the performance of neural networks and traditional learning models trained on hand-engineered statistical features for audio. Moreover, in our experiments, we show that given enough data, it is possible for deep neural networks to learn features directly on raw audio and attain classification accuracy comparable to the best performing traditional model learning from hand-engineered statistical features.