The audio + dense-caption mixture used to train
laion/voiceclap-small and
laion/voiceclap-large.
Each tar shard is a WebDataset of
paired
.flac (48 kHz mono audio) + .json (caption + metadata)
samples. Captions and structured attribute annotations are produced
automatically by a pipeline of audio-aware LLMs — Qwen-Audio, Gemini Flash 2.5,
and a thinking-mode reasoning model that scores emotion under the EmoNet
taxonomy plus per-clip vocal-burst, timbre… See the full description on the dataset page: https://huggingface.co/datasets/laion/voiceclap-data.
