Sketching the Expression: Flexible Rendering of Expressive Piano Performance with Self-Supervised Learning
- Paper: https://arxiv.org/abs/2208.14867
- Demo: https://free-pig-6c6.notion.site/DEMO-c20a1fea7a0844468a05b971c3b9ef3c
Clone this repository, and install the required packages:
git clone https://github.com/rsy1026/sketching_piano_expression.git
pip install -r requirements.txt
Then run this command to make sketching_piano_expression a python package for importing inner functions easily:
pip install -e .
We should align three files to parse the features: score MusicXML, score MIDI and human performance MIDI. Raw data samples can be found in ./scripts/data/data_samples/raw_samples.
Please make sure that MusicXML data has no error:
- Please check if any notes are hidden by some annotations (such as trills, glissandos, etc.) in the MusicXML score (You may check with the MuseScore software).
- Trills should not be abbreviated as "tr." signs but all notes should appear within each trill.
- Make sure that any complex techniques are not abbreviated as annotations but every single note should be written in the MusicXML score.
- Functions to extract the features may be updated in the future for imperfectly aligned MusicXML-MIDI files. These functions may flexibly skip the unaligned notes.
python3 extract_features.py --xml [filename1].musicxml --score [filename2].mid --perform [filename3].mid --measures [num1] [num2]
Enter to 'scripts' directory, and run the following command:
python3 ./data/make_batches.py --input_dir ./data/data_samples/features
You can make h5 files using the following function in make_batches.py:
create_h5_datasets(dataset='train', savepath='./data/data_samples')
Simply run the following command:
python3 train.py
You may change any settings within the code.