RELATE
收藏资源简介:
RELATE数据集是一个开源数据集,用于主观评估文本与音频之间的相关性。该数据集包含合成的音频样本和相关性评分,旨在帮助简化文本到音频(TTA)技术的评估过程。数据集涵盖了三个属性:听众、合成的音频和文本,并研究了这些属性对主观评价分数的影响。数据集由东京大学和庆应义塾大学的研究人员创建,并用于构建一个预测模型,该模型能够从合成的音频中自动预测主观评价分数。数据集包含9,963个评价、2,862个音频-文本对、28,806秒的音频时长和1,085名听众的评价。数据集的创建过程包括收集原始音频样本、合成音频样本、主观评价分数和听众属性。数据集的应用领域是文本到音频技术,旨在解决音频样本与输入文本内容的相关性问题。
The RELATE dataset is an open-source dataset for subjective evaluation of the correlation between text and audio. It contains synthesized audio samples and correlation scores, aiming to simplify the evaluation process of text-to-audio (TTA) technologies. The dataset covers three attributes: listeners, synthesized audio and text, and investigates the impact of these attributes on subjective evaluation scores. Developed by researchers from The University of Tokyo and Keio University, the dataset is utilized to build a predictive model that can automatically predict subjective evaluation scores from synthesized audio. It includes 9,963 evaluations, 2,862 audio-text pairs, 28,806 seconds of total audio duration, and ratings from 1,085 listeners. The dataset creation process involves collecting raw audio samples, synthesized audio samples, subjective evaluation scores and listener attributes. The dataset is targeted for text-to-audio technology applications, aiming to address the correlation issue between audio samples and input text content.




