Melody extraction based on source-filter modelling

This repository contains the code of the algorithm evaluated in MIREX 2015 and 2016 (BG). It also contains the code necessary to run the experiments in the following article (ISMIR2016):

J. J. Bosch, R. M. Bittner, J. Salamon, and E. Gómez, "A Comparison of Melody Extraction Methods Based on Source-Filter Modelling", in Proc. 17th International Society for Music Information Retrieval Conference (ISMIR 2016), New York City, USA, Aug. 2016.

J. Bosch, E. Gómez, "Melody extraction based on a source-filter model using pitch contour selection", in Proc. 13th Sound and Music Computing Conference (SMC 2016). Hamburg, Germany, 2016. p.67-74

Author: Juan J. Bosch Music Technology Group, Universitat Pompeu Fabra, Barcelona Contact: juan.bosch@upf.edu

This repository also contains code by R.M. Bittner (in contour_classification folder), and J.L Durrieu (source-filter model), which has been adapted to the needs of the conducted experiments

The code is written in python (version 2.7), and presents the following dependencies:

Essentia 2.0.1 or newer, with python bindings (http://essentia.upf.edu/) NumPy 1.8.2 (any relatively recent version should work)

For contour classification, the following packages are also used:

pandas scipy seaborn sklearn

In order to execute the algorithm evaluated in MIREX 2016 (BG1 and BG2 submissions), it should be called from the folder which contains the source code, as:

python MelodyExtractionFromSingleWav.py /inputaudiofolder/audio1.wav /estimations/audio1.txt --extractionMethod='BG2' --hopsize=0.01 --nb-iterations=30

In order to execute the algorithm based on energy weighting, it should be called from the folder which contains the source code, as:

python MelodyExtractionFromSingleWav.py /inputaudiofolder/audio1.wav /estimations/audio1.txt --extractionMethod='EWM' --hopsize=0.01 --nb-iterations=30

In order to execute the algorithm with the contour creation parameters from ISMIR2016, use --extractionMethod='CBM' :

python MelodyExtractionFromSingleWav.py /inputaudiofolder/audio1.wav /estimations/audio1.txt --extractionMethod='CBM' --hopsize=0.005805 --nb-iterations=30

Best results are generally obtained with a hopsize of 128 samples if sampling rate = 44100 (hopsize=0.0029025), but they take longer to compute:

python MelodyExtractionFromSingleWav.py /inputaudiofolder/audio1.wav /estimations/audio1.txt --extractionMethod='CBM' --hopsize=0.0029025 --nb-iterations=30

where %input is the path to a wav file, and output is the file with the estimated melody.

It is possible to also run the extraction using only Harmonic Summation instead of using source filter models. This option would be similar to the plugin MELODIA, but using the open source implementation in Essentia. This way, you can also save contours as those used in Bittner 2015 ISMIR article (and later use within a Pitch Contour Classification method) To do so use the option:

--extractionMethod='SAL'

To run contour classification experiments, you should first compute and save the contours, and adapt the paths accordingly. Make sure to use the same hopsize for contour creation and contour classification.

python run_contour_training_melody_extraction.py

python run_glass_ceiling_experiment.py