datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
voiceclap-data
VoiceCLAP Data
The audio + dense-caption mixture used to train
laion/voiceclap-small and
laion/voiceclap-large.
Each tar shard is a WebDataset of
paired <key>.flac (48 kHz mono audio) + <key>.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.voiceclap-flattened
voiceclap-flattened
SNAC-tokenized, flattened + augmented build of 5 subsets of
laion/voiceclap-data
(CC-BY-4.0): emolia (English block), ears, expresso, voxceleb1,
voxceleb2. Replaces the earlier single-subset
voiceclap-emolia-flattened repo (consolidated here).
Why this exists
Follow-up to an ablation study (2/3/5) on <listen>/<speak> token format:
full 7-tok/frame <listen> content with a SEPARATE vocab (ids shifted
+1,000,000 vs <speak>'s identical SNAC codes)… See the full description on the dataset page: https://huggingface.co/datasets/EmpathicRobotics/voiceclap-flattened.