Recurrent Neural Networks for Music Genre Classification

Music genre classification refers to identifying bits of music that belong to a certain tradition by assigning labels called genres. Recommendation systems automatically use classification techniques to group songs into their respective genres or to cluster music with similar genres. Studies show deep Recurrent Neural Networks (RNN) are capable of resolving complex temporal features of the audio signal and identifying music genres with good accuracy. This research experiments with different variants of RNN including LSTM, and IndRNN on the GTZAN dataset to predict the music genres. Scattering transforms along with Mel-Frequency Cepstral Coefficients (MFCCs) are used to construct the input feature vector. This study investigates various LSTM and simple RNN network architectures. Experiment results show a 5-layered stacked independent RNN was able to achieve 84% accuracy based on the aforementioned input feature vector.