If you are using this source code please consider citing the following paper:

D. Basaran, S. Essid and G. Peeters, Main Melody Extraction with Source-Filter NMF and CRNN, In 18th International Society for Music Information Retrieval Conference, ISMIR, 2018, Paris, France

Bibtex

	@inproceedings{basaran2018CRNN,
	Address = {Paris, France},
    	Author = {Basaran, D. and Essid, S. and Peeters, G.},
    	Booktitle = {19th Int.~Soc.~for Music Info.~Retrieval Conf.},
    	Month = {Sep.},
    	Title = {Main Melody Extraction with Source-Filter NMF and CRNN},
    	Year = {2018}
	}

To compute dominant melody estimation with the trained CRNN model, you can run the script

predict/predict_on_single_audio_CRNN.py

An example usage exists inside the script. An executable version of the prediction code is also available in the codeocean platform with the following link,

https://codeocean.com/2018/10/01/main-melody-extraction-with-source-filter-nmf-and-crnn/code

Feel free to play with it!

To create HF0 activation representation for a single track or the whole dataset, you can run the script

SF_NMF/extract_HF0.py

An example usage exists inside the script.

To create random train/validation/test splits, you can run the script

random_dataset_splits/random_splits.py

An example usage exists in the ReadMe.txt file. Note that random splitting requires HF0 representations, hence one has to first create HF0 representations then is able to use this script.

To train the model on the random splitted dataset, you can run the script

CRNN/C-RNN_model1.py

Note that if you want to use a GPU for the training part (probably you should), you would need to adjust the code for that purpose!

The required packages for the environment in the CRNN experiments are given in the requirements.txt file. Note that the main packages needed are

tensorflow_gpu, keras, pandas, numpy, scipy, scikit-learn, librosa, mir_eval, matplotlib, h5py,