Dataset Card for RepeatSynth

Synthetic portion of the audio datasets referenced in Class-Agnostic Audio Repetition Counting.

Code: https://github.com/Hamozwa/audio-counting

Dataset Summary

RS, RSN, and RVN are synthetic datasets containing varying levels of background noise. Every sample is uniformly 10 seconds long and contains between 0 and 8 repetition events.

Config	Description
rs	RepeatSound — clean, no added background noise
rsn	RepeatSoundNoisy — background noise added
rvn	RepeatVeryNoisy — the noisiest condition

Each config has train, validation, and test splits. rs-features, rsn-features, and rvn-features hold precomputed spectrograms and self-similarity matrices for the same clips (see "Feature configs" below) — most users only need the main configs above.

from datasets import load_dataset
ds = load_dataset("Hamozwa/RepeatSynth", "rsn", split="train")

Dataset Structure

Field	Type	Description	Why it's here
audio	Audio	The 10-second wav clip	The actual training/test input
id	string	Original filename stem	Lets you cross-reference against the -features configs or reconstruct the original file layout
original_source	string	The FSD50K clip this sample's repeated event was built from	Required for attribution — FSD50K licenses each clip individually, so this is what you check to find the license of the audio actually used in a given sample
num_repetitions	int	Ground-truth repetition count (0–8)	The label WavCounter and TSSMCounter are trained/evaluated against
events	list[{start: float, end: float}]	Onset/offset (seconds) of each repeated event within the clip	Used to build the target heatmap WavCounter regresses against, rather than just a single count — this is what lets the model localize events in time, not just count them

Feature configs

rs-features/rsn-features/rvn-features contain id, spectrogram, and tssm (temporal self-similarity matrix) columns: the precomputed inputs TSSMCounter and spectrogram baseline actually train on, joinable back to the main config via id. These are included for convenience and exact reproducibility, but are fully derivable from the audio column above via the preprocessing code in the linked repo, so they're kept as separate configs rather than bundled into the main table.

Dataset Sources

Synthetic clips were generated by combining two source collections:

• FSD50K — https://zenodo.org/records/4060432 — provided the audio for each repeated event
• MUSAN — http://openslr.org/ — used as background noise

Licensing

This is a derivative work built from two source collections, each with different terms. Both require attribution.

FSD50K licenses each clip individually rather than as a whole collection (CC0, CC-BY, CC-BY-NC, or CC Sampling+, depending on the clip). Since every sample here is built from one specific FSD50K clip, the original_source field lets you look up that clip's exact license via FSD50K's own dev_clips_info_FSD50K.json / eval_clips_info_FSD50K.json mapping files. If you plan to use or redistribute a subset commercially, check the license of every clip referenced by original_source first — some are CC-BY-NC and restrict commercial reuse.

MUSAN is released under CC BY 4.0 (https://openslr.org/17/) — attribution required, otherwise unrestricted.

Citation

[Your paper's BibTeX entry here]