WavReward: Spoken Dialogue Models With Generalist Reward Evaluators
Shengpeng Ji, Tianle Liang, Yangzhuo Li, Jialong Zuo, Minghui Fang, Jinzheng He, Yifu Chen, Zhengqing Liu, Ziyue Jiang, Xize Cheng, Siqi Zheng, Jin Xu, Junyang Lin, Zhou Zhao
2025-05-15
Summary
This paper talks about WavReward, a new system that uses AI to judge how well computer programs can hold spoken conversations, by listening to the audio and giving feedback based on how good the conversation is.
What's the problem?
The problem is that it's tough to measure how natural or helpful a computer's spoken conversation is, because most current methods don't really understand the deeper meaning or flow of a conversation, and they often just use simple scoring systems.
What's the solution?
The researchers created WavReward, which uses advanced audio language models to listen to conversations and then give more thoughtful and complex feedback, taking into account things like reasoning and the overall quality of the dialogue, not just basic features.
Why it matters?
This matters because it helps improve the way we train and test voice assistants and other spoken dialogue systems, making them sound more natural and useful for people in real-world situations.
Abstract
WavReward, a reward feedback model based on audio language models, evaluates the conversational performance of spoken dialogue systems, addressing the gap in evaluation by incorporating deep reasoning and nonlinear rewards.