KU-AGI/

LAION-COCO_dualcodec_pretokenize

DualCodec (12 Hz) pre-tokenized speech for omni speech–vision MLLM training. Source: LAION-COCO (top_caption); speech synthesized with CosyVoice.

	Contents (train split)




	




	

Total samples ≈ 500,000 (image–caption–speech triples)

Total audio ≈ 645 hours

WebDataset shards see repo (.tar)

Mean / median duration 4.64s / 4.33s

p90 / max duration 6.67s / 25.42s

Length cap 30s… See the full description on the dataset page: https://huggingface.co/datasets/KU-AGI/LAION-COCO_dualcodec_pretokenize.