CAF-Score：用于无参考音频标题评估的校准CLAP方法
声音
2026-03-23 v1 人工智能
计算与语言
摘要
尽管大型音频语言模型（LALMs）已推进音频标题生成，但鲁棒评估仍困难。基于参考的指标成本高昂且常常无法评估声学保真度，而基于CLAP（Contrastive Language-Audio Pretraining）的方法常常忽略语法错误和细粒度细节。我们提出CAF-Score，一种校准CLAP粗粒度语义对齐与LALMs细粒度理解和语法意识的无参考指标。通过将对比音频文本嵌入与LALM推理结合，CAF-Score有效检测语法不一致和细微幻觉。在BRACE基准测试上进行实验表明，我们的方法实现了最高的人类判断相关性，甚至在具有挑战性的情境下也优于基于参考的基线。这些结果凸显了CAF-Score对于无参考音频标题评估的有效性。代码和结果均可用于https://github.com/inseong00/CAF-Score。
引用
@article{arxiv.2603.19615,
title = {CAF-Score: Calibrating CLAP with LALMs for Reference-free Audio Captioning Evaluation},
author = {Insung Lee and Taeyoung Jeong and Haejun Yoo and Du-Seong Chang and Myoung-Wan Koo},
journal= {arXiv preprint arXiv:2603.19615},
year = {2026}
}
备注
A condensed version of this work has been submitted to Interspeech 2026. Section 10 is an extended analysis added in this version