Music Genre Classification with Parallel Neural Networks

Part of the ECE 542 Virtual Symposium (Spring 2020)

Music streaming services such as Spotify, Pandora, and SoundCloud serve their customers by providing access to literally millions of songs. With such a vast number, it is important that these companies be able to classify songs by their characteristics for retrieval by their customers. In particular, users should be able to search by genre. To facilitate such a feature, these companies need an automated method to classify the genre of the songs in their databases. Our approach uses various parallel network architectures – two different networks trained individually on the same data similarly to Music Genre Classification with Paralleling Recurrent Convolutional Neural Network, Leng et. al. – and compares performances of these parallel networks to a simple fully connected network. A convolutional neural network can extract meaningful representations from image data and a recurrent neural network can learn discriminative features from sequential data. With this in mind, we train both networks on the Short-time Fourier Transform spectrogram representation of each, pop off the classification layer of each network to get the learned representation from each network, concatenate those outputs, and feed the concatenation into a perceptron for classification. This approach achieves 86% genre classification accuracy – more than 15% above human recognition.