Viewer note: default uses viewer_preview/ for responsive audio playback. Full training/evaluation files remain available in the original folder structure.

MAESTRO-E

MAESTRO-E dataset for music practice error detection with paired performance/reference inputs and note-level error labels.

Paired Inputs for Error Detection

The model takes paired inputs:

• mistake: performance audio/MIDI containing musical errors
• score: paired reference score audio/MIDI (target/correct context)

Error supervision is provided with labels:

• label/extra_notes
• label/removed_notes
• label/correct_notes

Full Structure (Code-Compatible)

This repo preserves the filesystem layout expected by the training loader:

• mistake/<sample_id>/...
• score/<sample_id>/...
• label/{extra_notes,removed_notes,correct_notes}/<sample_id>/...
• maestro-v3.0.0.json (included at repo root)

Quick Start (Directly with Existing Code)

from huggingface_hub import snapshot_download

root = snapshot_download("ben2002chou/maestro-e", repo_type="dataset")
# set root_dir = root
# set split_json_path = f"{root}/maestro-v3.0.0.json"

Project Link

• https://github.com/ben2002chou/LadderSYM

Citation

@inproceedings{
chou2026laddersym,
title={LadderSym: A Multimodal Interleaved Transformer for Music Practice Error Detection},
author={Benjamin Shiue-Hal Chou and Purvish Jajal and Nicholas John Eliopoulos and James C. Davis and George K Thiruvathukal and Kristen Yeon-Ji Yun and Yung-Hsiang Lu},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=cizuvfyQXs}
}