Score Transformer

This is the official page for “Score Transformer: Generating Musical Score from Note-level Representation” (ACM Multimedia Asia 2021).

Paper

Code

Poster

Overview

Musical scores contain various musical elements, which help us visually comprehend music.

Generating musical scores from MIDI needs to estimate a variety of musical elements that are not represented in MIDI.

Score Tokenization

We design score token representation to handle musical scores with transformers.

Tokenization tools are available on our repository!

Training Scheme

We train Transformer model to translate MIDI into musical scores.

This scheme requires musical scores only.

Example

Trained model succeeds in generating appropriate musical scores out of MIDI.

More examples to come!

Performance

Differences between the original and generated scores were measured based on a metric.

Score Transformer vs. Baselines

Score Transformer works significantly better than baselines (including Finale and MuseScore).

Score Token vs. Existing score formats

Score token outperforms existing score formats (ABC, Humdurm, and LilyPond).

Citation

If you find our work is helpful, please consider citing our work:

@inproceedings{suzuki2021,
 author = {Suzuki, Masahiro},
 title = {Score Transformer: Generating Musical Score from Note-level Representation},
 booktitle = {Proceedings of the 3rd ACM International Conference on Multimedia in Asia},
 year = {2021},
 pages = {31:1--31:7},
 doi = {10.1145/3469877.3490612}
}