SingMOS
收藏资源简介:
SingMOS是由中国人民大学信息学院和卡内基梅隆大学语言技术研究所联合创建的高质量歌唱语音数据集,旨在解决歌唱领域中MOS评价数据稀缺的问题。该数据集包含3421个中文和日文歌唱片段,总时长4.25小时,平均长度4.47秒。数据集内容丰富,涵盖了21种歌唱语音合成模型、6种歌唱语音转换模型和6种再合成模型。创建过程中,使用了多种开源工具和模型,确保了数据的多样性和可靠性。SingMOS数据集的应用领域主要集中在歌唱MOS预测,为提升歌唱语音质量提供了重要数据支持。
SingMOS is a high-quality singing speech dataset jointly created by the School of Information, Renmin University of China and the Language Technology Institute, Carnegie Mellon University, aiming to address the scarcity of MOS evaluation data in the singing domain. This dataset contains 3,421 Chinese and Japanese singing speech clips, with a total duration of 4.25 hours and an average length of 4.47 seconds. Boasting rich content, the dataset covers 21 singing speech synthesis models, 6 singing voice conversion models and 6 resynthesis models. During its creation, various open-source tools and models were employed to ensure the diversity and reliability of the dataset. The application scenarios of SingMOS mainly focus on singing MOS prediction, providing critical data support for improving the quality of singing speech.




