Today in the development of voice AI, it faces a big obstacle, called a “linguistic ecosystem imbalance” because the amount of minor language data is very scarce, which makes the model performance difficult, and making it unfair. Grannary: It is Nvidia’s open-sourced speech dataset project announced on August 2025 It has collected around 1 million hours of people’s voice audio. NVIDIA Granary is the first industrial scale speech dataset to cover many minor European languages. It will thus be a landmark for the task. This article seeks to comprehensively understand the development of Granary by researching and comparing versions of the granary: Scale up, language up, quality up, and up to ethics. In this paper, not only to fill the gap on the research about the evolutionary trend of a single dataset, but also to get a concrete industrial-level data building pattern; In the future, it can provide very valuable advice to how to build an even more inclusive, robust and trusted multilingual speech intelligence.
Jiongzeng Ye (Mon,) studied this question.