Skip to the content.

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.

ScoreToken Tokenization tools are available on our repository!

Training Scheme

We train Transformer model to translate MIDI into musical scores.

Overview This scheme requires musical scores only.

Example

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

Result 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). Eval1

Score Token vs. Existing score formats

Score token outperforms existing score formats (ABC, Humdurm, and LilyPond). Eval2

Citation

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

@inproceedings{suzuki2021mmasia,
 author = {Suzuki, Masahiro},
 title = {Score Transformer: Generating Musical Score from Note-level Representation},
 journal = {ACM Multimedia Asia (MMAsia'21)},
 month = December,
 year = {2021},
 doi = {10.1145/3469877.3490612}
}