TTA-Bench
收藏资源简介:
TTA-Bench是一个用于评估文本到音频(TTA)模型的全面基准数据集,涵盖了功能性、可靠性和社会责任性等方面。数据集包含2999个通过自动化和手动方法生成的多样化提示,旨在对TTA模型进行全面评估。数据集的构建考虑了准确性、效率、泛化、鲁棒性、公平性、偏见和毒性等七个维度,并通过结合客观指标和超过118,000个来自专家和普通用户的注释来进行评估。
TTA-Bench is a comprehensive benchmark dataset for evaluating Text-to-Audio (TTA) models, covering aspects such as functionality, reliability, and social responsibility. It contains 2999 diverse prompts generated via both automated and manual approaches, aiming to conduct comprehensive evaluations of TTA models. The dataset was constructed with seven core dimensions in mind: accuracy, efficiency, generalization, robustness, fairness, bias, and toxicity, and its evaluation is carried out by combining objective metrics with over 118,000 annotations from both experts and general users.




