AuralReason

AuralReason provides bilingual speech with concise chain-of-thought supervision and assistant responses for voice-sensitive dialogue and speech emotion recognition.

Paper · Code and training · AURAL-RL

Configurations

Configuration	Examples	Languages	Task
LIME	314,855	Chinese, English	Voice-sensitive dialogue
EmotionCoT	17,840	English	Speech emotion recognition
HumanSpeech	74,330	Chinese, English	Voice-sensitive dialogue
Total	407,025

These are the complete filtered examples with newly constructed reasoning and responses. Each configuration provides a train split. EmotionCoT uses IEMOCAP Sessions 1–4 and MELD train/dev, together with Expresso, MSP-Podcast, and EARS.

GeneralSpeech is an in-house dataset and is not publicly released at this time.

Format

Audio is embedded in Parquet as 16 kHz mono FLAC. Each example has eight columns:

Column	Description
id	Stable sample identifier
audio	Audio bytes and a public FLAC filename
instruction	Task instruction, empty for audio-only dialogue
transcript	Spoken content
reasoning	Concise chain of thought
response	Assistant response or emotion label
language	zh or en
source	Original speech corpus

Load

import io
import soundfile as sf
from datasets import Audio, load_dataset

data = load_dataset("YuxiangW/AuralReason", "EmotionCoT", split="train")
data = data.cast_column("audio", Audio(decode=False))
row = data[0]
waveform, sample_rate = sf.read(io.BytesIO(row["audio"]["bytes"]))
print(row["reasoning"])
print(row["response"])

Select LIME or HumanSpeech to load the other configurations. The AURAL code repository includes a command that downloads the data and prepares local training JSONL and audio:

python -m aural.prepare_data --output data/auralreason

Sources

LIME combines its Chinese and English dialogue audio with ECD-TSE and ESD. EmotionCoT combines MELD, IEMOCAP, Expresso, MSP-Podcast, and EARS. HumanSpeech includes WenetSpeech, GigaSpeech, affective media, EmotionTalk, VoxMovies, RealMAN, ESD, CSEMOTIONS, and Hongloumeng. The source column retains the original corpus for each example.

The released CoT annotations and responses use Apache-2.0. Source recordings retain their original corpus licenses and attribution.

Citation

@article{wang2026aural,
  title={AURAL: Adaptive Latent Reasoning with Joint Chunk for Speech Language Models},
  author={Wang, Yuxiang and others},
  journal={arXiv preprint arXiv:2610.01560},
  year={2026}
}