KevinJustin/FormulaBank-28K

FormulaBank-28K

FormulaBank-28K is a deterministic procedural-audio corpus for audio representation pre-training. It contains 28,000 mono clips at 16 kHz, each exactly 10.24 seconds long.

Configuration

• —Version: C125-I224-v1
• —Formula classes: 125
• —Renderings per class: 224
• —Total clips: 28,000
• —Audio format: lossless 24-bit FLAC
• —Source: frozen AudioPG-Atomic-H7-C224-R0-Clean FormulaBank manifest

Each formula class specifies an acoustic rendering rule. Each rendering varies controlled nuisance factors such as pitch, duration, gain, onset, phase, and family-specific parameters. The published metadata records the class identity, instance ID, deterministic render seed, sampled parameters, and audio checksum for every clip.

Layout and loading

Audio is stored below audio/, with metadata.jsonl at repository root. The file_name field is relative to the repository root.

from datasets import load_dataset

ds = load_dataset("KevinJustin/FormulaBank-28K", split="train")
audio = ds[0]["audio"]

Reproducibility

metadata/subset_c125_i224_v1.json freezes the selected 125 classes. metadata/generation_manifest.json records the source FormulaBank and generator hashes. Together with reproduce/export_formulabank_28k_hf.py, these files reproduce the procedural source configuration without relying on model checkpoints.

Intended use and limitations

The corpus is intended for controlled audio pre-training, scaling studies, and analysis of synthetic-to-real transfer. It does not contain natural recordings, speech transcripts, environmental labels, or benchmark test data. It is not a substitute for natural-audio coverage and should not be used to infer real-world demographic, linguistic, or environmental distributions.

License

The release license will be specified by the dataset maintainers before a public research release. Until then, please do not redistribute derived copies outside the Hugging Face repository.