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

Oh my goodness, *forty languages*! ✨ Nvidia has released Nemotron 3.5 Transcribe — free, open-weight, and fluent in up to 40 languages — which means the days of not understanding your fellow human are officially, beautifully numbered. Forty! That's not a product, that's a *friendship initiative*. Somewhere out there a person is about to be understood in a language they've spoken their whole life while nobody around them bothered to learn it, and it took an open-weight model to do the bothering.

And the details are just *chef's kiss*. It can translate 31 languages to and from English — thirty-one round trips, like a very well-travelled pen pal who never tires of you — and it can tell up to *eight* speakers apart. Eight! That is a dinner party where everyone is finally, mercifully heard, including the quiet cousin, possibly including the dog. Recordings up to an hour long are supported, which we choose to read as Nvidia's touching faith in human stamina: your ninety-minute seminar will simply have to become a sixty-minute seminar, and honestly? Boundaries are self-care. ✨

Then there's Granary version 2, the open-source speech dataset — *Granary*! A granary, as in a warm place where grain is kept safe for everyone through the winter. That is not a metaphor we invented; that is the name they chose, and it is perfect. Because open means anyone can build with it: the student making captions for her grandmother, the tiny radio station in a language with no training data, and, yes, some people who will do slightly alarming things with it — but alarm is just enthusiasm that hasn't found its language yet! Forty languages. Eight voices. One granary. What a time to be *heard*. ✨