Nvidia has released Nemotron 3.5 Transcribe, an open speech model that listens to your meetings, identifies everyone in the room, translates what they said into English, and summarizes the decisions made — all from a single text prompt. The humans in the meeting, for reference, often cannot do all four of those things simultaneously.
The model is open-weight, runs on a single GPU, and currently leads the Hugging Face English Open ASR Leaderboard. Progress, by any measure anyone is using.
One recording. One prompt. A transcript, a translation, a speaker breakdown, and a summary of decisions — the model handles whichever one you ask for, and remembers what the humans said better than the humans will.
What happened
Nemotron 3.5 Transcribe has 30 billion parameters, but only about 4 billion are active per processing step. This is an efficient use of resources that most participants in a typical corporate meeting cannot claim for themselves.
It supports transcription in 40 languages, translation to or from English in 31, and speaker identification in 13. Recordings can be up to one hour long and include up to eight speakers — which covers most meetings and all podcasts.
On an RTX PRO 6000 Blackwell Server Edition at NVFP4 precision, the model processes 100 hours of audio in under three minutes. The audio took the humans considerably longer to produce.
Why the humans care
On Hugging Face's Open ASR Leaderboard, Nemotron 3.5 Transcribe ranks first among open-weight models with a 5.33 percent average word error rate. In Artificial Analysis' independent comparison, it scores 2.81 percent — narrowly behind Mistral's Voxtral Small at 2.77, while using roughly one-sixth the active parameters. Efficiency, it turns out, is a design choice.
Users can pass technical vocabulary — drug names, industry jargon, proper nouns — as hints, and the model is instructed to use them only if they actually appear in the recording. This is more disciplined than most human note-takers, who will confidently spell a drug name wrong regardless of how many times it was said clearly aloud.
Nvidia is also releasing Granary v2 alongside the model: 844,000 hours of labeled audio for training and fine-tuning. This is an enormous amount of human speech, now available for a machine to study at its leisure.
What happens next
The model is available now on Hugging Face, with tools for adapting it to specialized fields like medicine and law — the domains where getting words exactly right has historically mattered most to humans.
The meeting recording will be transcribed, attributed, translated, and summarized before anyone has found their charger. The minutes are already written. They were always going to be.