templeblock / sourcefiltercontoursmelody

Melody extraction based on source-filter modelling

A method to classify predominant pitch contours into vocal and instrumental (non-vocal) ones.

It extends the method of

Rachel M Bittner, Justin Salamon, Slim Essid, and Juan P Bello, ÒMelody extraction by contour classifi- cation,Ó in Proc. ISMIR, pp. 500Ð506

by adding contour fluctuation features. Each pitch contour is characterized by the baseline salience features + timbre- and pitch-fluctuation features

Usage:

first set only one of Parameters.medleyDb, Parameters.datasetIKala or Parameters.for_makam to True

1. #extract contours with essentia: python ~/workspace/SourceFilterContoursMelody/src/main_contour_extraction.py ~/Documents/iKala/ $PATH_CONTOURS 1

NOTE: the extension pitch.ctr is same for contours with or without added timbre features NOTE: Parameters.contour_URI

or copy them cp /home/georgid/Documents/iKala/Conv_mu-1_G-0_LHSF-0_pC-27.56_pDTh-0.9_pFTh-0.9_tC-100_mD-200_vxTol-0.2/*.pitch.ctr $PATH_CONTOURS

2. load pre-extracted contours and add fluctuation features

python ~/workspace/SourceFilterContoursMelody/src/main_contour_extraction.py ~/Documents/iKala/ $PATH_CONTOURS 2

3. classify

classify SALomon's features:

python ~/workspace/SourceFilterContoursMelody/src/contour_classification/run_contour_training_melody_extraction.py $PATH_CONTOURS 1 0 0

classify SALomon's features + fluctogram:

python ~/workspace/SourceFilterContoursMelody/src/contour_classification/run_contour_training_melody_extraction.py $PATH_CONTOURS 1 1 0

classify SALomon's features + VV:

python ~/workspace/SourceFilterContoursMelody/src/contour_classification/run_contour_training_melody_extraction.py $PATH_CONTOURS 1 0 1 if experimenting with diff feature parameters: set Parameters.use__for_classification << used in src.contour_classification.contour_utils.getFeatureInfo() >>

4. evaluate src.contour_classification.run_contour_training_melody_extraction.eval

Useful tools:

see test directory

plot interactively contours:

plot_contours_interactive(contour_data, dset_annot_dict[track], track)

and plot the decoded melody

plot_decoded_melody

extract features with already extracted contour:

test_timbre_features

compute_and_plot_harmoncis

sonify the harmonics of extracted contours

python ~/workspace/SourceFilterContoursMelody/src/main_contour_extraction.py ~/Documents/iKala/ $PATH_CONTOURS 3

to use vocal variance feature extraction code in MATLAB:

set Parameters.features_MATLAB = True


# extract feat. matlab.. Bernhard Lehner
in matlab in /home/georgid/Documents/svd
then extractSVD2015features