This is the source code of DeepChoir, a chord-conditioned melody harmonization system with controllable harmonicity, trained/validated on Chordified JSB Chorales Dataset.
The evaluation data we used in our experiments in the outputs folder, and the online discrimination test is available at https://sander-wood.github.io/deepchoir/test.
You can listening generated samples (chorale, folk, pop, and symphony) at https://sander-wood.github.io/deepchoir/samples.
In addition, we created an album, "Choirish", a collection of Irish folk music harmonized by DeepChoir. It is available at https://soundcloud.com/sander-wood/sets/choirish-a-collection-of-irish-folk-music-harmonized-by-deepchoir.
For more information, see our paper: arXiv paper.
As we did not find any full JSB Chorales dataset with human-annotated chord symbols, and as performing harmonic analysis manually is too time-consuming, we carried out the following automated preprocessing.
1. Chordifying: simplify a complex score with multiple voices into a succession of chords in one voice via a tool in music21.
2. Labelling: rank all chords based on beat strength, then only keep those with the highest scores and add them to sopranos.
We ended up with a total of 366 chorales for training (90%) and validation (10%). Admittedly, these automatically labelled chord symbols are not authoritative from the perspective of harmonic analysis, but they are sufficient for our task.
You can find this chordified version of JSB Chorales dataset in the dataset folder.
Chordified BWV 322 rendered by MuseScore3
Python: 3.7.9
Keras: 2.3.0
tensorflow-gpu: 2.2.0
music21: 6.7.1
tqdm: 4.62.3
samplings: 0.1.7
PS: Third party libraries can be installed using the pip install command.
1. Put the melodies with chord symbols (in the format of .musicxml, .xml or .mxl) in the inputs folder;
2. Simply run choralizer.py;
3. Wait a while and the harmonized melodies will be exported in the outputs folder.
You can set two parameters HARMONICITY∈[0, 1] in config.py to adjust two attributes of the generated chorales.
The higher the value of HARMONICITY, the generated three parts are more in harmony with the melody.
1. Store all the chorales with chord symbols in the dataset folder;
2. Run loader.py, which will generate corpus.bin;
3. Run model.py, which will generate weights.hdf5.
After that, you can use choralizer.py to generate chorales that fit the musical style of the new dataset.
If you need to finetune the parameters, you can do so in config.py. It is not recommended to change the parameters in other files.