SECD — String Ensemble Chord Dataset: Experiments

This repository provides the public GitHub companion materials for the SECD benchmark experiments, including Mini-SECD, benchmark notebooks, metrics, figures, and lightweight reproducibility/demo materials; the official full SECD dataset release is archived on Zenodo at https://doi.org/10.5281/zenodo.15547207, and the accompanying SECD paper citation remains pending publication.

---

Repository Structure

SECD/
├── dataset/                        ← Dataset description & access info
├── mini_secd_demo/                 ← GitHub-friendly demo dataset
├── experiments/
│   ├── EXP1_ensemble_size/         ← Classify duo / trio / quartet
│   ├── EXP2_chord_quality/         ← Classify chord quality
│   ├── EXP3_dynamics/              ← Classify dynamics
│   └── EXP4_tec_fam/              ← Classify playing technique family
└── requirements.txt

---

Experiments Summary

Exp	Task	Classes	Test Acc	Test Macro F1
EXP1	Ensemble size	duo / trio / quartet	98.67%	98.64%
EXP2	Chord quality	major / minor / diminished / augmented	93.73%	93.73%
EXP3	Dynamics classification (arco, per-instrument)	pp / p / mp / f / ff	98.19%	98.01%
EXP4	Technique family	bowed / col_legno / harmonic / plucked	99.39%	97.29%

---

Setup

git clone https://github.com/ageroul/SECD.git
cd SECD
pip install -r requirements.txt

The experiments are provided as Jupyter/Colab notebooks. They require the Python dependencies in requirements.txt, access to the AST backbone weights, and either the included Mini-SECD demo dataset or the full SECD dataset.

---

Mini-SECD Demo Dataset

This repository includes a small demo dataset at mini_secd_demo/. Mini-SECD is for notebook execution and pipeline validation only. It is not a reproducibility dataset and should not be used to report or compare scientific metrics.

Mini-SECD contains real precomputed SECD mel-spectrogram tensors (.npy) copied from the full SECD mel cache. The WAV files in mini_secd_demo/ are tiny valid silent placeholders; they exist only because the original notebooks check for audio-path existence before loading cached mel tensors. They are not the source audio used in the paper.

The CSV metadata in Mini-SECD is reconstructed from mel filenames and SECD schema conventions for demo execution. It is not copied from the original experiment CSV metadata. Use Mini-SECD to verify that data loading, filtering, splitting, label extraction, dataset construction, and short demo-mode execution work end to end.

---

Reproducibility

The metrics shown above come from the full experimental setup. Reproducing those results requires the official full SECD release on Zenodo:

https://doi.org/10.5281/zenodo.15547207

Full metric reproduction also requires:

• the full SECD dataset,
• the full precomputed SECD mel cache,
• the original/full metadata,
• the required Python dependencies,
• access to the pretrained AST backbone weights,
• and a suitable GPU runtime.

The notebooks preserve the original experiment definitions: tasks, labels, model architectures, losses, and evaluation metrics are unchanged. Demo mode changes only runtime settings such as dataset root, device selection, dataloader workers, batch size, and the number of training steps. These changes make the notebooks easier to execute on normal machines, but they do not turn Mini-SECD into a paper-results reproduction package.

---

Quick Start

Install dependencies, then open any notebook under experiments/.

pip install -r requirements.txt

By default, when mini_secd_demo/ is present, notebooks run in demo mode:

RUN_MODE = "demo"
SECD_BASE = "./mini_secd_demo"
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

Demo mode is suitable for checking that the notebooks execute and that the pipeline is wired correctly. It is not suitable for reproducing or validating the reported paper scores.

To use the full dataset instead, set SECD_RUN_MODE=full and point SECD_BASE at the full SECD dataset root before running the notebooks:

export SECD_RUN_MODE=full
export SECD_BASE=/path/to/full/SECD

CPU execution is supported for demo runs. Full training remains GPU-oriented and can be slow or impractical on CPU.

---

Model Backbone

All experiments use the Audio Spectrogram Transformer (AST) pretrained on AudioSet.

---

License and Citation

The GitHub code and documentation in this repository are released under the MIT License.

Mini-SECD is provided only as a lightweight demo and pipeline-validation package. It is not the official full SECD dataset and is not intended for reproducing the reported benchmark metrics.

If you use SECD, Mini-SECD, the benchmark notebooks, or these GitHub materials in academic work, please cite the official SECD Zenodo DOI and/or the accompanying paper once available. The official full SECD dataset release is archived on Zenodo:

https://doi.org/10.5281/zenodo.15547207

---

Citation

@misc{secd_dataset_2026,
  title = {SECD: String Ensemble Chord Dataset},
  note  = {Official full dataset release archived on Zenodo. Paper citation pending publication.},
  year  = {2026},
  doi   = {10.5281/zenodo.15547207},
  url   = {https://doi.org/10.5281/zenodo.15547207}
}