ASLP-lab/Emo-Emilia

Emo-Emilia Dataset

To better simulate real-world context, we introduce a new SER test set, Emo-Emilia. Specifically, we apply the automated labeling approach to annotate Emilia, a large-scale multilingual and diverse speech generation resource with over 100,000 hours of speech data that captures a wide range of emotional contexts. We then manually verify the accuracy of the emotion labels. Each utterance is checked by at least two experts to ensure both accuracy and reliability. The final proposed test set, Emo-Emilia, consists of 1400 test samples, with 100 samples per emotion category across seven types (angry, happy, fearful, surprised, neutral, sad and disgusted) in both Chinese and English (700 samples per language).

Emo-Emilia is a subset of Emilia dataset. The original Emilia dataset can be accessed here.

You can download the Emo-Emilia data file on HuggingFace here. More audio information can be found in the ./Emo-Emilia/Emo-Emilia-ALL.jsonl file

For more information, please refer to our paper"Steering Language Model to Stable Speech Emotion Recognition via Contextual Perception and Chain of Thought"