Benchmarking Audio-Based Empathetic Response Judgment in Large Audio Language Models
Hosted by NLP2CT Lab, University of Macau · View source on GitHub
Context: "Okay. Um, you don't sound very much like you think that's gonna work."
Speaker: " Yeah. It's, you know, it's one thing to control my own drinking, and it's another thing when everyone else around me is drinking and then there's like that pressure -of, "Oh, they're gonna think I look stupid.""
Listener: "Okay. So, how are you gonna quit drinking without your friends thinking you look stupid?"
Annotations: Emotional Reaction: 0 Exploration: 1 Interpretation: 1 Genuineness: 1 Prosody: 1 Warmth: 1
Context: "You wanna write everybody's story for them?"
Speaker: "I really think I'd be a shitty writer. I'd want to give everybody a happy ending."
Listener: "I'm very sorry about your patient."
Annotations: Emotional Reaction: 2 Exploration: 0 Interpretation: 2 Genuineness: 2 Prosody: 2 Warmth: 2
Context: "Did you call the police?"
Speaker: "No but if it happens again I will. I ordered a camera for my porch. I couldn't get any sleep that night."
Listener: "Do you live in a house or an apartment?"
Annotations: Emotional Reaction: 0 Exploration: 1 Interpretation: 0 Genuineness: 1 Prosody: 1 Warmth: 0
The demo website source and the audio examples shown here are publicly available in the NLP2CT repository. The full benchmark, prompts, and evaluation scripts are not included in this demo repository.
📦 Download Demo (ZIP)
💻 GitHub Repository
The website code is open source under the MIT License. Audio samples and other dataset content retain their original terms; the code license does not relicense them.