configs:
- config_name: single_turn
data_files:
- split: train
path: single_turn/train/*.tar
- config_name: multi_turn
data_files:
- split: train
path: multi_turn/train/*.tar
license: cc-by-4.0
language:
- en
size_categories:
- 10M<n<100M
EchoDialogue
EchoDialogue is a large-scale English dataset for empathetic spoken dialogue, combining emotionally expressive speech, fine-grained acoustic annotations, and structured cognitive reasoning. It supports research on understanding users’ emotions and psychological needs, selecting appropriate support strategies, and generating empathetic responses in both content and vocal expression.
The dataset contains 360K+ single-turn examples and 4K+ five-turn conversations. Each example includes user and response audio, corresponding text, emotion and speaking-style annotations, and structured reasoning traces.
Key features include:
- Rich emotional expression: Diverse emotional categories and intensities, including challenging cases such as sarcasm and conflicting verbal and vocal cues.
- Fine-grained acoustic supervision: Annotations of pitch, energy, speaking rate, intonation, word stress, and emotional intensity.
- Structured empathetic reasoning: A three-stage process connecting emotional and acoustic perception, mental-state inference, and empathetic strategy selection.
- Diverse dialogue settings: Single-turn and multi-turn conversations cover varied emotional expressions, sarcasm, and evolving conversational states.
If you find EchoDialogue useful for your research, we would appreciate an acknowledgment and link to this dataset page. The accompanying paper and its BibTex citation will be added once available.