LibriS2S

LibriS2S是由马斯特里赫特大学数据科学与知识工程系创建的德英双语语音对齐数据集，旨在解决语音到语音翻译领域的数据稀缺问题。该数据集包含12本独立录制的德语和英语有声书音频，通过精确的语音对齐技术，确保了两种语言发音的无偏性。数据集的创建过程涉及从Librivox项目中提取音频，并使用aeneas工具进行文本与音频的对齐。LibriS2S的应用领域主要集中在语音合成和语音到语音翻译模型的训练，特别是在FastSpeech 2模型的改进中，该数据集展示了其在提升语音合成质量方面的潜力。

LibriS2S is a German-English bilingual speech-aligned dataset created by the Department of Data Science and Knowledge Engineering, Maastricht University, aiming to address the data scarcity issue in the speech-to-speech translation domain. This dataset includes audio recordings of 12 independently recorded German and English audiobooks. Through precise speech alignment techniques, it ensures the unbiasedness of pronunciation for both languages. The dataset creation process involves extracting audio from the Librivox project and aligning text with audio using the aeneas tool. The main application scenarios of LibriS2S focus on the training of speech synthesis and speech-to-speech translation models. Particularly in the improvement of the FastSpeech 2 model, this dataset has demonstrated its potential in enhancing the quality of speech synthesis.