NVIDIA open-sources Nemotron 3.5 Transcribe and the 463,000-hour Granary v2 speech dataset

NVIDIA's speech release is promptable and adaptable. It handles transcription, diarization of up to 8 speakers, translation across 31 languages and audio question answering, with a 64k context window. It ships under the OpenMDW-1.1 license.

Granary v2 is arguably the bigger gift: a 463,062-hour curated public speech corpus covering 48 languages.

The model is also coming to RTX Spark laptops. It competes with Mistral's Voxtral, which ranks #2 on Voice Arena's diarization leaderboard.

What to watch: independent word-error-rate results on low-resource languages.