Heroes

Heroes数据集是由名古屋大学和Meta AI共同创建的，专注于电视系列领域的表达性语音到语音翻译测试集。该数据集包含406个样本，每个样本平均时长为3.10分钟，主要用于评估语音翻译系统在维持翻译准确性的同时，如何有效地转移源语音的韵律属性，如语调、强调和情感。数据集的创建过程包括去噪、质量检查和人工转录校正，确保数据质量。该数据集的应用领域主要集中在提高语音到语音翻译系统的表达性，解决跨语言交流中的韵律信息传递问题。

The Heroes dataset was co-created by Nagoya University and Meta AI, and it is an expressive speech-to-speech translation test set focused on the television series domain. It contains 406 samples, with an average duration of 3.10 minutes per sample. It is primarily used to evaluate how effectively speech translation systems can transfer the prosodic attributes of source speech, such as intonation, emphasis and emotion, while maintaining translation accuracy. The dataset's creation process includes denoising, quality inspection and manual transcription correction to ensure data quality. Its main application areas focus on enhancing the expressiveness of speech-to-speech translation systems and addressing the issue of prosodic information transfer in cross-linguistic communication.