WavCaps

dataset·weak audio captions·active

Large weakly labeled audio-caption corpus assembled from web sound libraries and AudioSet, with ChatGPT-assisted caption cleaning.

Recorded facts

Official site	https://huggingface.co/datasets/cvssp/WavCaps ↗
Geography	United Kingdom
arxiv	2303.17395
github repository	XinhaoMei/WavCaps
size	audio clips: 403050 · average duration seconds: 67.59
scope	Audio-language pretraining across general audio, including music-relevant clips.
genres	mixed audio; includes music
license	CC BY 4.0 dataset card
steward	University of Surrey CVSSP and collaborators
creators	Xinhao Mei; Chutong Meng; Haohe Liu; Qiuqiang Kong; Tom Ko; Chengqi Zhao; Mark D. Plumbley; Yuexian Zou; Wenwu Wang
languages	English
modalities	FLAC; JSON captions; raw descriptions; metadata
geographies	Global
access method	Hugging Face metadata and waveform archives plus GitHub code.
intended uses	audio captioning; audio-language pretraining; retrieval
related papers	WavCaps: A ChatGPT-Assisted Weakly-Labelled Audio Captioning Dataset
annotation method	Three-stage filtering and ChatGPT-assisted transformation of noisy source descriptions.
collection method	Audio sourced from FreeSound, BBC Sound Effects, SoundBible and AudioSet Strong.
provenance claims	Dataset card identifies four upstream sources but does not unify their underlying content licenses.
current availability	Hugging Face dataset card and downloads available.
commercial use limits	Dataset card does not state a noncommercial limit; upstream assets may differ.
us market scope basis	Publicly available or materially used in US-facing music/audio-AI research.
limitations biases disputes	Captions are weak labels transformed by a language model.; Source licenses and quality vary across four collections.
redistribution restrictions	Must reconcile upstream audio terms in addition to dataset-card attribution.