LibriVoxDeEn

LibriVoxDeEn是由海德堡大学计算语言学与IWR创建的一个大型德语到英语语音翻译和德语语音识别数据集。该数据集包含110小时的德语音频，对应超过50,000条平行句对，以及547小时的德语语音用于语音识别。数据来源于德语有声书，确保了音频质量高且流畅度低。创建过程中，使用了先进的音频-文本和文本-文本对齐工具，并通过人工评估确保了高对齐质量。该数据集主要应用于解决德语到英语的直接语音翻译问题，支持端到端神经语音翻译模型的训练。LibriVoxDeEn is a large-scale German-to-English speech translation and German speech recognition dataset created by the Department of Computational Linguistics and IWR, Heidelberg University. This dataset includes 110 hours of German audio paired with over 50,000 parallel sentence pairs, alongside 547 hours of German speech for speech recognition tasks. The data is sourced from German audiobooks, which ensures high audio quality and low disfluency. During the dataset construction process, advanced audio-text and text-text alignment tools were utilized, and manual evaluation was carried out to guarantee high alignment quality. This dataset is primarily used to address direct German-to-English speech translation tasks, supporting the training of end-to-end neural speech translation models.