Comparison of music genre classification with spectrogram input vs. raw audio input using 1D & 2D Convolutional Neural Networks. In this experiment, only the first 20 seconds of each audio is used. Each audio samples is split into 10 parts of 2 seconds audio.

- Python 2
- Numpy
- Matplotlib
- Scikit-learn
- Scikit-plot
- Keras
- Tensorflow
- Kapre
- Librosa
- ffmpeg

GTZAN music genre data set by George Tzanetakis. The dataset consists of 1000 audio tracks each 30 seconds long. It contains 10 genres, each represented by 100 tracks.

	Confusion Matrix	ROC curve	Test accuracy
Raw audio input with 1D CNN			0.31
Spectrogram input with 1D CNN			0.7372
Spectrogram input with 2D CNN			0.686

• Dieleman, Sander, and Benjamin Schrauwen. "End-to-end learning for music audio." Acoustics, Speech and Signal Processing (ICASSP), 2014 IEEE International Conference on. IEEE, 2014.
• https://github.com/tuwien-musicir/DL_MIR_Tutorial
• https://github.com/Hguimaraes/gtzan.keras