CREMA-ASIS

CREMA-ASIS is a dataset of 77,559 synthetic English utterances in which the acoustic emotion (how something is said) and the semantic sentiment (what is said) are controlled independently. It covers both incongruous pairs, such as an angry voice saying something positive, and congruous ones.

It accompanies the paper When Vocal Tone and Literal Meaning Diverge: An Acoustic–Semantic Incongruity Study for Large Audio–Language Models (Findings of EMNLP 2026). Code for generation, filtering, fine-tuning and evaluation is on GitHub.

Splits follow the original CREMA-D speaker splits, so no actor appears in more than one split.

Split	Samples	Actors
train	65,534	63
val	6,146	14
test	5,879	14

Usage

from datasets import load_dataset

ds = load_dataset("WilliamHo1999/CREMA-ASIS", split="test")
sample = ds[0]
sample["audio"], sample["acoustic"], sample["semantic"]

Limitations and ethical considerations

The speech is synthetic and imitates the voices of real CREMA-D actors. Do not use it to impersonate those individuals. Emotion labels are the intended emotion of the reference clip; the AER filter removes clear mismatches but does not guarantee every sample conveys its label.

License

CREMA-ASIS is derived from CREMA-D and is released under the Open Database License (ODbL) v1.0, the same license as CREMA-D. The text comes from GoEmotions (Apache 2.0).

Citation

If you find our paper helpful, please consider citing our papers and starring us！

@article{chen2026vocal,
  title={When Vocal Tone and Literal Meaning Diverge: An Acoustic-Semantic Incongruity Study for Large Audio-Language Models},
  author={Chen, Yu-Wen and Ho, William and Topaz, Maxim and Kostic, Zoran and Hirschberg, Julia},
  journal={arXiv preprint arXiv:2608.28966},
  year={2026}
}