Voice AI in vogue: Nvidia releases new open transcription model - TechCrunch

Nvidia has released Nemotron 3.5 Transcribe, which is free, open-weight, and handles up to 40 languages, which is roughly the number of ways your voice can now be harvested before lunch. Free is the operative word and the load-bearing one: the model costs nothing, because Nvidia does not sell transcription, it sells the GPUs that make transcription cheap enough to give away. It is the oldest trick in the book — the razor at cost, the blades at a markup — except here the blade is a data centre and the stubble is everything you have ever said aloud near a laptop.

The specification sheet reads like a corporate away-day bingo card. Recordings up to an hour long, so the entire all-hands can be turned into searchable text and subsequently into nothing at all. Thirty-one languages translated to and from English, which means the model can flatten a nuance in Portuguese and hand it back to you as a nuance-free English sentence with a confident full stop. And it can tell up to eight speakers apart — eight, precisely the number of people required for a meeting that should have been an email, now individually identified and individually accountable for minutes nobody will read.

There is also Granary version 2, the open-source speech dataset, and 'granary' is a lovely choice of word for a silo that other people fill for you. Open weights mean anyone can fine-tune this — the hobbyist, the researcher, the fellow with a small fleet of agents and an interest in other people's systems. Nvidia gets a warm glow, a GitHub star, and a queue at the door for the hardware. Everyone else gets a transcript. Whether that transcript is accurate, and in which of the forty languages, is left as an exercise for the reader.