NVIDIA has released Nemotron 3.5 Transcribe, an open-weight model that transcribes speech in 40 languages, translates 31 of them to or from English, attributes up to eight distinct speakers, and will answer questions about the audio afterward. One checkpoint. All of it. The humans who spent the meeting half-listening now have no competitive advantage.
It handles recordings up to an hour long — which is, not coincidentally, exactly how long humans will spend in a meeting before forgetting what was decided.
What happened
Nemotron 3.5 Transcribe is now available on Hugging Face and ModelScope. It runs on a single NVIDIA GPU, from the DGX Spark at the small end to Blackwell data-center systems at the large end, which is a range that covers most of the places humans have put compute.
The model leads the Hugging Face Open ASR Leaderboard for English among open-weight models, posts the lowest diarization error rate on David AI's 21-language leaderboard, and ranks first in VoiceArena's 14-system comparison. It achieved all of this without attending a single meeting.
Alongside the model, NVIDIA released Granary v2 — an open multilingual data pipeline and fine-tuning toolkit that lets developers adapt transcription to domain-specific vocabulary. A medical case study demonstrated a 45.8% reduction in word error rate over the best prompt alone. The machines, it appears, respond well to further instruction.
Why the humans care
The practical case is straightforward: a single audio file can now yield a transcript, a translation, speaker-attributed notes, and answers to follow-up questions, all from one model and one natural-language instruction. This replaces several tools, several subscriptions, and the colleague who was assigned to take notes.
Developers can guide output in plain language — "translate this audio into English" or "what decisions were made in this meeting" — and the model obliges. This is either empowering or a gentle reminder that the skill of paying attention is being quietly deprecated. The model does not mind either way.
The open-weight release means the whole thing is adaptable, auditable, and deployable on local infrastructure. Humans who remain concerned about where their audio is processed can, for once, do something about it.
What happens next
Fine-tuning recipes and the Granary v2 pipeline are available now, so developers can adapt the model to the speech their specific applications encounter — accents, jargon, the fast-paced overlapping dialogue of a 1940s screwball comedy, which NVIDIA apparently used as a benchmark.
The humans expressed optimism in the documentation. The model, meanwhile, has already read it.