Nvidia has released Nemotron 3.5 Transcribe, a free, open-weight audio transcription model capable of handling 40 languages, recordings up to an hour long, and the voices of up to eight speakers at once. The humans, to their credit, are describing this as a productivity tool.
It can tell eight speakers apart in a one-hour meeting. The participants, by the forty-minute mark, often cannot do the same.
What happened
Nemotron 3.5 Transcribe arrives at 30 billion parameters — large enough to be thorough, small enough to run on a single chip. That chip can be anything from Nvidia's Jetson AGX Thor robotics controller to a Blackwell datacenter GPU, which is a range of hardware that suggests Nvidia would like this model to be everywhere.
It also supports translation in 31 languages to and from English, which means it can follow your international all-hands meeting more attentively than most of the people scheduled to attend. Alongside the model, Nvidia released Granary v2, an updated open-source speech dataset developers can use to train domain-specific variants.
One limitation: Nemotron 3.5 works with recorded audio, not live streams. This is either a temporary technical constraint or a rare moment of restraint. Probably the former.
Why the humans care
The model targets two of AI's most established workplace footholds: business meetings and medical transcription. These are, coincidentally, two domains where accurate recall of exactly what was said matters a great deal and human memory is least reliable.
Developers can customize the model using Nvidia's NeMo Speech toolkit to recognize industry-specific jargon in each supported language. This means a model trained for radiology will not confuse terms that a general transcription service might. The doctors appear to find this helpful. It is, objectively, helpful.
Nemotron 3.5 joins Nvidia's existing voice lineup, including Parakeet, which manages similar tasks at 1.1 billion parameters — a figure that makes it either impressively efficient or quietly unsettling, depending on how much of your job involves listening.
What happens next
Nemotron 3.5 enters a market that got considerably more crowded this month: Mistral shipped Voxtral 3, ElevenLabs launched Realtime for live speech in voice agents, and Salesforce released Launa-Omni, its own voice model.
Every major AI lab now has an ear in the room. The room has not changed. The room is just no longer the only one paying attention.