Speech-DRAME
收藏资源简介:
Speech-DRAME是一个统一的框架,旨在解决语音角色扮演的评估问题。它提供了三个层面的贡献:一个具有双语人工注释数据的评估基准,一个经过微调的评估模型,以及一个语音角色扮演基准。Speech-DRAME区分了两种互补的评估策略:原型评估和现实主义评估。与零样本ALLM评估器相比,DRAME-Eval与人类评分的吻合度更高。通过整合透明的基准资源、建模方法和系统级评估,Speech-DRAME为评估语音提供了第一个全面、可重复的基础。
Speech-DRAME is a unified framework designed to tackle the evaluation challenge of speech role-playing tasks. It provides three core contributions: a bilingual human-annotated evaluation benchmark, a fine-tuned evaluation model, and a dedicated speech role-playing benchmark. Speech-DRAME differentiates two complementary evaluation strategies: prototype evaluation and realism evaluation. DRAME-Eval demonstrates a significantly higher correlation with human ratings compared to zero-shot ALLM evaluators. By integrating transparent benchmark resources, modeling methods and system-level evaluation, Speech-DRAME establishes the first comprehensive and reproducible foundation for speech role-playing evaluation.




