quinnlue/FSD50K-16k

FSD50K 16 kHz FLAC

A clean, streamable repack of Fhrozen/FSD50k, with all 51,197 clips converted from 44.1 kHz PCM WAV to lossless mono 16 kHz, 16-bit FLAC and embedded in Hugging Face-native Parquet shards.

The official development split is exposed as train and validation; the official evaluation set is exposed as test. Freesound IDs, weak multi-label annotations, AudioSet MIDs, uploader metadata, per-clip licenses, tags, descriptions, and PP/PNP human ratings are preserved.

Dataset summary

Split	Clips	Hours	Parquet shards
train	36,796	70.39	13
validation	4,170	10.35	2
test	10,231	27.85	5

Total duration: 108.59 hours.

Usage

from datasets import load_dataset

ds = load_dataset("quinnlue/FSD50K-16k", streaming=True)
sample = next(iter(ds["train"]))
print(sample["audio"]["sampling_rate"])  # 16000
print(sample["labels"])

Schema

• audio: embedded FLAC audio, mono, 16 kHz, 16-bit
• freesound_id, freesound_url: stable source identity and source page
• labels, mids: smeared FSD50K labels and matching AudioSet ontology identifiers
• title, description, tags, uploader: uploader-provided Freesound metadata
• license: the Creative Commons URL for the individual clip
• human_ratings: available PP/PNP ratings grouped by AudioSet MID
• duration_seconds: source clip duration
• source_partition: original dev or eval partition

Original annotation and metadata files are retained under metadata/ for reproducibility. The exact source revision is ccf1acaff12f3f4a4c10052dddafbb6d22152c9b. checksums.json contains SHA-256 hashes for every data shard.

Conversion

Audio was converted with FFmpeg using:

-ac 1 -ar 16000 -sample_fmt s16 -c:a flac -compression_level 5

Conversion is lossless with respect to the resampled 16-bit PCM signal; resampling from 44.1 kHz to 16 kHz is not reversible. No loudness normalization, clipping, trimming, or content filtering was applied.

Licensing and attribution

FSD50K as a curated dataset is released under CC BY 4.0. Individual clips carry their own Creative Commons license (including CC0, CC BY, CC BY-NC, and Sampling+); consult each row's license, uploader, freesound_id, and freesound_url before reuse. This repack does not replace or relax those clip-level terms.

Please cite the original work:

@article{fonseca2020fsd50k,
  title={FSD50K: An Open Dataset of Human-Labeled Sound Events},
  author={Fonseca, Eduardo and Favory, Xavier and Pons, Jordi and Font, Frederic and Serra, Xavier},
  journal={arXiv preprint arXiv:2010.00475},
  year={2020}
}