Music Genre Classification with RNNs

• Classify music files based on genre from the GTZAN music corpus
• GTZAN corpus is included for easy of use
• Use multiple layers of (bidirectional) Recurrent Neural Nets
• Implementations in PyTorch and Keras.

Test trained RNN model

In the ./weights/ you can find trained model weights and model architecture.

To test the model on your custom audio file, run

 python3 predict_example.py path/to/custom/file.mp3

or to test the model on our custom files, run

 python3 predict_example.py audios/classical_music.mp3

Audio features extracted

• MFCC
• Spectral Centroid
• Chroma
• Spectral Contrast

Dependencies

• Keras or PyTorch
• numpy
• librosa - for audio feature extraction

Ideas for improving accuracy:

• Normalize MFCCs & other input features (Recurrent BatchNorm?)
• Decay learning rate
• How are we initing the weights?
• Better optimization hyperparameters (too little dropout)
• Do you have avoidable bias? How's your variance?

Accuracy

• Training (at Epoch 400): Training loss: 0.5801 Training accuracy: 0.7810

• Validating: Dev loss: 0.734523485104 Dev accuracy: 0.766666688025

• Testing: Test loss: 0.900845060746 Test accuracy: 0.683333342274