Reproduce “SDiaReward: Modeling and Benchmarking Spoken Dialogue Rewards with Modality and Colloquialness” on OpenResearch

About this paper

Jingyu Lu, Yuhan Wang, Fan Zhuo, Xize Cheng, Changhao Pan, Xueyi Pu, Yifu Chen, Chenyuhao Wen and 2 more

SDiaReward introduces an end-to-end multimodal reward model and a new benchmark, ESDR-Bench, for evaluating spoken dialogue systems by explicitly addressing the modality and colloquialness gaps. The model, trained on large-scale preference data, achieves 96.61% accuracy in distinguishing human speech from synthetic audio and demonstrates robust generalization across diverse conversational domains and unseen TTS engines.

Key insights

• Targeted pairwise supervision is essential for learning subtle paralinguistic preferences in spoken dialogue, as general-purpose audio judges prioritize semantic content over acoustic naturalness.
• The modality gap is a more significant challenge for evaluation than the colloquialness gap; while linguistic cues can infer 'spoken style,' distinguishing audio with identical text but different prosody requires specialized modeling.
• The proposed reward model implicitly learns a relative ranking function calibrated to the style and difficulty of different conversational domains (e.g., scripted vs. wild dialogues), indicating an understanding of relative expressiveness.

Summary by alphaXiv.