Merkel Podcast Corpus

Merkel Podcast Corpus是一个多模态数据集，由汉堡大学等机构的研究人员基于德国前总理安格拉·默克尔的16年每周视频播客创建。该数据集包含音频、视觉和文本三种模态，涵盖了默克尔的演讲和访谈，总时长达到48小时。创建过程中，研究人员通过下载视频、转录文本和其他元数据，使用强制对齐和说话人识别技术来精炼数据集。该数据集主要用于解决单说话人深度学习模型的训练问题，同时也适用于多说话人模型的泛化，以及跨语言处理任务，如唇同步配音。

Merkel Podcast Corpus is a multimodal dataset created by researchers from institutions including the University of Hamburg based on the 16-year weekly video podcasts of former German Chancellor Angela Merkel. The dataset contains three modalities: audio, visual and text, covering Merkel's speeches and interviews, with a total duration of 48 hours. During the dataset creation process, researchers downloaded the videos, transcribed the textual content and collected other metadata, and used forced alignment and speaker recognition technologies to refine the dataset. This dataset is mainly used for training single-speaker deep learning models, and is also applicable to the generalization of multi-speaker models and cross-language tasks such as lip-sync dubbing.