VoxPopuli

VoxPopuli是由Facebook AI创建的大型多语种语音数据集，包含23种语言共计40万小时的未标记语音数据，是目前最大的开放数据集，用于无监督表示学习和半监督学习。该数据集还包含15种语言的1.8K小时转录演讲及其对15种目标语言的口头解释，总计17.3K小时。VoxPopuli旨在通过提供丰富的多语种音频数据，推动多语种自动语音识别（ASR）和语音翻译（ST）的研究进展，解决现有数据集在多语种支持上的不足，并特别关注实时语音翻译（解释）的质量与延迟平衡问题。

VoxPopuli is a large-scale multilingual speech dataset developed by Facebook AI. It contains 400,000 hours of unlabeled speech data spanning 23 languages, making it the largest open dataset currently available for unsupervised representation learning and semi-supervised learning. Additionally, the dataset includes 1.8K hours of transcribed speeches in 15 languages, paired with verbal explanations for 15 target languages, totaling 17.3K hours. VoxPopuli aims to advance research in multilingual automatic speech recognition (ASR) and speech translation (ST) by providing abundant multilingual audio data, addressing the limitations of existing datasets in multilingual support, with a particular focus on balancing quality and latency in real-time speech translation (interpretation).