LadderSym

Music Error Detection with Iterative Inter-Stream Alignment and Symbolic Score Prompting

LadderSym is a transformer-based system for end-to-end music performance error detection.

Setup

Clone the repository and set up the environment:

git clone https://github.com/ben2002chou/LadderSYM.git
cd LadderSYM
conda create -n laddersym python=3.11 -y
conda activate laddersym
pip install -r requirements.txt
wandb login

Download datasets and official checkpoints (recommended):

python setup_laddersym_assets.py --datasets all --checkpoints all
source .env.laddersym

Manual setup is also supported (see Data and Checkpoints below).

Training

python train_laddersym.py \
  --config-name=config_maestro \
  model=laddersym_MT3Net \
  dataset=MAESTRO \
  split_frame_length=2000

python train_laddersym.py \
  --config-name=config_coco \
  model=laddersym_MT3Net \
  dataset=CocoChorales \
  split_frame_length=2000

Evaluation

MAESTRO-E:

python test_laddersym.py \
  --config-dir=config \
  --config-name=config_maestro \
  model=laddersym_MT3Net \
  path=/path/to/model.ckpt \
  eval.eval_dataset=MAESTRO \
  eval.exp_tag_name=laddersym_maestro \
  hydra/job_logging=disabled \
  eval.contiguous_inference=True \
  split_frame_length=2000

CocoChorales-E:

python test_laddersym_coco.py \
  --config-dir=config \
  --config-name=config_coco \
  model=laddersym_MT3Net \
  path=/path/to/model.ckpt \
  eval.eval_dataset=CocoChorales \
  eval.exp_tag_name=laddersym_coco \
  hydra/job_logging=disabled \
  eval.contiguous_inference=True \
  split_frame_length=2000

Inference

Batch inference over piece folders (recommended):

python laddersym_test_inference.py \
  --config-dir=config \
  --config-name=config_maestro_prompted \
  model=laddersym_MT3Net \
  path=pretrained/laddersym/checkpoints/maestro/prompted/model.ckpt \
  dataset_dir=$LADDERSYM_MAESTRO_ROOT \
  output_mid_name=laddersym_output.mid \
  overwrite=true \
  hydra/job_logging=disabled

Single-piece inference with explicit file paths:

python laddersym_test_inference.py \
  --config-dir=config \
  --config-name=config_maestro_prompted \
  model=laddersym_MT3Net \
  path=pretrained/laddersym/checkpoints/maestro/prompted/model.ckpt \
  mistake_file=$LADDERSYM_MAESTRO_ROOT/<piece_id>/mistake.wav \
  score_file=$LADDERSYM_MAESTRO_ROOT/<piece_id>/score.wav \
  prompt_file=$LADDERSYM_MAESTRO_ROOT/<piece_id>/score.mid \
  output_dir=$LADDERSYM_MAESTRO_ROOT/<piece_id> \
  output_mid_name=laddersym_output.mid \
  hydra/job_logging=disabled

Interpreting Outputs

The output MIDI contains three semantic tracks:

• Track 1: Extra notes
• Track 2: Missing notes
• Track 3: Correct notes
• Example output path: <piece_dir>/laddersym_output.mid

Data and Checkpoints

• Datasets: CocoChorales-E, MAESTRO-E
• Checkpoints repo: ben2002chou/laddersym-checkpoints
• Viewer-friendly previews: CocoChorales-E-preview, MAESTRO-E-preview

Recommended setup:

python setup_laddersym_assets.py --datasets all --checkpoints all
source .env.laddersym

Checkpoint variants:

Dataset	Prompted	Unprompted
CocoChorales-E	model.ckpt	model.ckpt
MAESTRO-E	model.ckpt	model.ckpt

Advanced setup options:

# only datasets
python setup_laddersym_assets.py --datasets all --checkpoints none

# only checkpoints
python setup_laddersym_assets.py --datasets none --checkpoints all

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}
}