BeepBank-500 (earcons-mini-500)

BeepBank-500 is a compact, fully synthetic earcon/alert mini‑dataset (≈300–500 clips) for UI sound research. It contains short tones and triads generated from a controlled parameter grid (waveform family, f0, duration, envelope, amplitude modulation, and simple Schroeder-style reverbs). The dataset ships with a metadata schema, lightweight baselines, and a data note template for arXiv. Audio is intended for release under CC0-1.0 (public domain). Code is MIT-licensed.

Why this dataset? Reproducible, rights‑clean, and tiny enough to bundle in a 2–3 day sprint as a citable asset. Typical tasks: earcon classification, timbre analysis, f0 regression, onset detection, psychoacoustic proxies.

Quick start (2–3 day plan)

0) Create environment

python -m venv .venv && source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt

1) Generate audio + metadata (≈400 clips by default)

python code/generate_earcons.py --outdir audio --meta metadata/metadata.csv --seed 13 --target_n 400

Optional: change --target_n to 500 for a bigger set.

2) (Optional) Recompute/augment features later

python code/compute_features.py --indir audio --meta_in metadata/metadata.csv --meta_out metadata/metadata.csv

3) Run tiny baselines

# Waveform family classifier (sine/square/triangle/fm_*)
python code/baselines/classify_waveform.py --audio_dir audio --meta metadata/metadata.csv

# f0 regression MAE (Hz) on single‑tone items only
python code/baselines/f0_regression.py --audio_dir audio --meta metadata/metadata.csv

4) Archive and publish

• Create a Zenodo record (v1.0.0) with /audio, /metadata, /code, README.md, LICENSE, CITATION.cff.
• Mirror code on GitHub; connect Zenodo badge; tag a release.
• Submit the arXiv data note using docs/data_note_arxiv.tex (subject: eess.AS). Link the Zenodo DOI.

5) Data Load

from datasets import load_dataset ds = load_dataset("mandip/beepbank-500")

Folder layout

earcons-mini-500/
├─ audio/                    # WAV files (mono, 48kHz, 16-bit PCM)
├─ metadata/
│  ├─ metadata.csv           # one row per file with parameters and features
│  └─ LICENSES.md            # audio license statement (CC0-1.0)
├─ code/
│  ├─ generate_earcons.py    # main generator (no external DSP deps)
│  ├─ compute_features.py    # recompute features if needed
│  └─ baselines/
│     ├─ classify_waveform.py
│     └─ f0_regression.py
├─ docs/
│  ├─ data_note_arxiv.tex    # 2–3k word data note skeleton
│  └─ figures/
├─ requirements.txt
├─ CITATION.cff
├─ LICENSE                   # MIT for code
├─ CHANGELOG.md
└─ README.md

Dataset design notes

• Waveforms: sine, square, triangle, FM (2:1 & 3:2), optional chords (major/minor triads).
• Durations: 100/250/500 ms; Envelopes: fast, medium, percussive.
• AM: none, 8 Hz (0.3), 30 Hz (0.5).
• Reverbs: simple Schroeder-style small (0.3 s) and medium (0.6 s).
• Normalization: RMS target with peak cap at −1 dBFS. (If pyloudnorm is installed, LUFS can be computed; normalization remains RMS‑based by default.)
• Metadata: includes generation params plus features (peak/rms dBFS, centroid, bandwidth, zcr, proxies).
• Splits: deterministic hash-based (train/val/test).

Licensing

• Audio: CC0‑1.0 (public domain dedication). Keep attribution unnecessary.
• Code: MIT License (see LICENSE).
• If you later add CC‑BY assets, list them in metadata/LICENSES.md with full attribution and URLs.

How to cite

Zenodo DOI : "https://doi.org/10.5281/zenodo.17172015". arxiv : "https://arxiv.org/abs/arXiv:2509.17277"

---

Maintainer: Mandip Goswami
Scope: niche psychoacoustic/UI earcon research
Keywords: earcon, alarm, psychoacoustics, timbre, AM, ADSR, reverb