Paralinguistic Speech Captions (ParaSpeechCaps)
收藏资源简介:
Paralinguistic Speech Captions (ParaSpeechCaps)是一个覆盖了59种独特风格标签的大型数据集,由德克萨斯大学奥斯汀分校和纽约大学共同创建。该数据集包括342小时的人工标注数据(PSC-Base)和2427小时的自动标注数据(PSC-Scaled)。数据集旨在通过丰富的风格标签来提升文本到语音合成的风格表现,包含说话人级别的内在标签和语句级别的情境标签。数据集创建过程中,首先通过众包方式收集了282小时语音的内在大规模标注数据,然后提出两种新颖的数据扩展方法,一种针对内在标签,一种针对情境标签,以自动扩展数据集规模。该数据集可应用于风格化文本到语音合成,解决语音合成中的风格控制问题。
Paralinguistic Speech Captions (ParaSpeechCaps) is a large-scale dataset covering 59 unique style labels, co-created by The University of Texas at Austin and New York University. It includes 342 hours of manually annotated data (PSC-Base) and 2427 hours of automatically annotated data (PSC-Scaled). This dataset aims to improve the stylistic performance of text-to-speech synthesis through rich style labels, and contains speaker-level intrinsic tags and utterance-level contextual tags. During the dataset construction, 282 hours of intrinsic large-scale annotated speech data were first collected via crowdsourcing, then two novel data expansion methods were proposed, one targeting intrinsic tags and the other targeting contextual tags, to automatically scale up the dataset size. This dataset can be applied to stylized text-to-speech synthesis, addressing the style control problem in speech synthesis.

- 1Scaling Rich Style-Prompted Text-to-Speech Datasets德克萨斯大学奥斯汀分校计算机科学系, 纽约大学计算机科学与数据科学系 · 2025年



