To Trust or Not to Trust: Retrieval-Augmented Fact Checking in Speech

AI 导读

文章提出VeriSpeak基准测试，揭示语音事实验证中文本与语音模态差异，证明检索结合推理可提升准确性至86.1%。

VeriSpeak benchmark highlights modality gap in speech fact-checking, showing retrieval combined with reasoning achieves 86.1% accuracy.

重点速览

• 创建VeriSpeak基准测试评估语音事实验证能力 Introduces VeriSpeak benchmark for speech fact verification
• 发现文本与语音模态存在显著验证差异 Identifies modality gap between text and speech verification
• 检索增强模型常混淆证据与语音内容 Retrieval alone struggles with evidence-claim distinction
• 推理结合检索可提升事实核查准确率 Reasoning combined with retrieval boosts accuracy
• 数据集公开支持研究复现与应用 Dataset available for research and application

一句话：语音事实核查需语音理解与证据推理双重能力 / Speech fact-checking requires both audio understanding and evidence reasoning.

推荐理由 为AI语音验证提供跨模态研究范式，具实际应用指导价值