This is a written project with regard to processing sound using keras deep learning library. These are the planned application:
- Genres classification
- Music auto-tagging
- Instrumental classification
- Tempo classification
- Deep Learning Jazz generator
- Sort music by genres, artist, tempo, tags
- Piano analysis
In this small project, I try to make a classifier with two dimensional neural network to distinguish among 11 genres of music. The data I use contains 1000 pieces of 30 seconds music that are separated into different classes from GTZAN, and it can be downloaded here or on the GTZAN website.
- model.h5 file is a weight of 1D Convolutional network with shape (max_feature, max_channels). Feature train: stft.
- model2D.h5 file is a weight of 2D Convolutional network with shape (max_feature, time_length, max_channels). Did not uploaded due to large file.
- model2D_2.h5 file is also a weight of 2D Convolution network, with the same shape but the channel is set in last instead of front.
- Conclusion: Feature engineering is not good due to low accuracy, more have to be done with extracting feature from data!
- After extracting the data as spectrogram and create model based on 128x1308 pixel, I successfully train the model with an accuracy of 99.90 % and loss of 1.61%. I realize that reducing the learning rate to the factor of 1e(+n) may improve the chance of weights being rounded errorly.
- Plan for the next part: write a clear data generator for the network, write cleaner code, write documentation, seperate files by utilities, write test.
- Update Training Analysis
