YODAS v3：百万小时开放语音数据集发布
原标题：YODAS v3: A 1 Million Hour Dataset for the Next Generation of Open Voice AI Research
AI 摘要
Hugging Face Blog 发布 YODAS v3，一个 110 万小时的多语言语音数据集，规模是 YODAS v1 的三倍，覆盖 100 多种语言。该数据集全部为 48kHz，超过 92% 的数据有效采样率达到 32kHz 以上，且 70% 以上包含至少两个不同声道，并附带词级和话语级时间戳及部分英文翻译。它采用 CC BY 3.0 许可，旨在为开放语音 AI 研究提供大规模、高质量、多通道的训练资源，支持 ASR、语音合成、空间音频生成和语音翻译等任务。
正文节选
It has become quite clear that training data is the true kingmaker in today's era of AI research. Architectures come and go, but the lesson of the last several years has been consistent: models get better when they see more data, and more varied data. Nowhere is this more true than in voice. Speech is one of the richest signals humans produce. It carries not just words but also accent, emotion, speaking style, background noise, and the texture of real conversation. The systems that have reached