Modeling and Benchmarking Spoken Dialogue Rewards with Modality and Colloquialness
New reward model SDiaReward evaluates full speech conversations, not just text, achieving state-of-the-art preference accuracy.
A research team has introduced SDiaReward, a novel AI model designed to evaluate the quality of spoken dialogue systems by addressing two critical shortcomings in current methods. The model tackles the 'modality gap'—assessing prosody and emotion in speech that text alone cannot capture—and the 'colloquialness gap,' which distinguishes polished written scripts from the disfluencies and spontaneity of natural conversation. Trained on a new dataset of episode-level preference pairs, SDiaReward operates directly on full multi-turn speech episodes, providing a holistic assessment of conversational flow and expressiveness.
To validate its performance, the team also established ESDR-Bench, a stratified benchmark for robust, episode-level evaluation. Experiments show SDiaReward achieves state-of-the-art pairwise preference accuracy, significantly outperforming general-purpose audio large language models (LLMs). The analysis suggests the model captures deeper conversational expressiveness beyond superficial audio cues, leading to improved generalization across different domains and recording conditions. The release includes code, data, and demos, providing a comprehensive toolkit for developers to build more natural and engaging voice-based AI agents.
- SDiaReward model evaluates full multi-turn speech episodes for prosody, emotion, and naturalness.
- It addresses the 'modality gap' and 'colloquialness gap' where current text-based methods fail.
- Paired with the new ESDR-Bench, it significantly outperforms general audio LLMs on preference accuracy.
Why It Matters
Enables development of more natural, expressive voice assistants and conversational AI by providing a better training signal.