Piano Difficulty Estimator from Audio

This demo is based on our paper accepted at TASLP 2024:
https://ieeexplore.ieee.org/document/10878288

It predicts how difficult a piano piece is to play, using only audio (MP3 or YouTube).
Unlike other systems that use scores or images, this model works directly from sound.

We introduce PSyllabus, a dataset with 7,901 labeled piano pieces in 11 difficulty levels.
The models support CQT, pianoroll, and multimodal inputs.

This demo is intended for researchers in music education and musicology.

Note: the demo is very slow on CPU. We kindly request a small GPU.
It runs smoothly on a laptop with an RTX 3060 (6 GB).