SyMuPe: MIDI Quality Classifier

MIDI-Quality-Classifier is a model trained to automatically assess the quality of symbolic piano performances. It classifies MIDI files into four distinct categories: score (inexpressive/rendered), high quality, low quality, and corrupted.

Introduced in the paper: PianoCoRe: Combined and Refined Piano MIDI Dataset.

TISMIR: https://doi.org/10.5334/tismir.333
arXiv: https://arxiv.org/abs/2605.06627
SyMuPe: https://github.com/ilya16/SyMuPe
PianoCoRe: https://github.com/ilya16/PianoCoRe
Dataset: https://huggingface.co/datasets/SyMuPe/PianoCoRe

Architecture

Type: Transformer Encoder
Backbone: 12-layer Transformer (80M parameters) pre-trained on the deduped subset of the Aria-MIDI dataset using a Multi-Mask Language Modeling (mMLM) objective.
Classification Module: Single layer transformer and a classification head.
Objective: Sequence Classification (4 classes).
Inputs (score-agnostic): Pitch, Velocity, TimeShift, Duration, absolute TimePosition
Classes:
- Score (S): Rendered or synthesized scores with constant tempo/dynamics.
- High Quality (HQ): Clean expressive performances (recorded or high-fidelity transcriptions).
- Low Quality (LQ): Transcriptions with noticeable noise or minor errors.
- Corrupted (C): Broken files or severely failed transcriptions.
Training: Trained for 20,000 iterations on created subset of the PianoCoRe dataset as described in the paper.

Quick Start

Before using this model, ensure you have the symupe library installed:

pip install -U symupe

Use the following code to classify MIDI files:

import torch
from symupe import AutoClassifier

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Build Classifier by loading the model and tokenizer directly from the Hub
classifier = AutoClassifier.from_pretrained(
    "SyMuPe/MIDI-Quality-Classifier", device=device
)
# model, tokenizer, labels = classifier.model, classifier.tokenizer, classifier.labels

# Classify a MIDI file
result = classifier("performance.mid")
# result is MusicClassificationResult(...) containing:
# - midi, seq, probabilities, prediction, label, all_logits, all_probabilities, all_predictions,
#   sequences and window_indices

print(f"Predicted Label: {result.label}")
print(f"Probabilities: {result.probabilities}")

License

The model weights are distributed under the CC-BY-NC-SA 4.0 license.

Citation

If you use this model or the associated dataset in your research, please cite:

@inproceedings{borovik2025symupe,
  title = {{SyMuPe: Affective and Controllable Symbolic Music Performance}},
  author = {Borovik, Ilya and Gavrilev, Dmitrii and Viro, Vladimir},
  year = {2025},
  booktitle = {Proceedings of the 33rd ACM International Conference on Multimedia},
  pages = {10699--10708},
  doi = {10.1145/3746027.3755871}
}

@article{borovik2026pianocore,
  title = {{PianoCoRe: Combined and Refined Piano MIDI Dataset}},
  author = {Borovik, Ilya},
  year = {2026},
  journal = {Transactions of the International Society for Music Information Retrieval},
  volume = {9},
  number = {1},
  pages = {144--163},
  doi = {10.5334/tismir.333}
}