Voxtral 3 is now on tabi
Mistral shipped new voice models, and tabi got an upgrade.
By Nikhil Bhima, founder of tabi
The launch film, 37 seconds. Best with sound.
Until this week, tabi ran on Mistral’s earlier voice models. They were good, and every feature in tabi was built on them.
On 20 October, Mistral released Voxtral 3, its new family of speech models. Within a day, tabi was running on it. Every take you say is now heard by Voxtral 3, and everything tabi reads to you is spoken by it.
Here’s what that changed.
It hears you better
Voxtral 3 has sharper ears. On my own test recordings, it heard twice as many whispers word for word as the old model. Names I’d taught tabi came out right every time, 12 out of 12, where the old model got 4.
56 languages, up from 13
You can now dictate in 56 languages, including Marathi, Gujarati, Punjabi, Urdu, Persian and Tagalog. Pick yours in Settings and tabi tells the model what to expect.
New voices, and Listen in 37 languages
Read aloud now uses Voxtral 3’s voice, with 36 new voices to pick from. Listen, the page in tabi for PDFs, books and photos of pages, now reads and translates into 37 languages, up from 9.
Names, said your way
Type a name in the Dictionary, press Say it, and say it once. tabi learns how it hears the name and how you pronounce it, so it’s typed right and read aloud the way you say it.
What I fixed along the way
Ears this sharp can pick up words in a quiet room. So a short take now also goes to the earlier model, which hears silence as silence. If it hears nothing, tabi types nothing.
A name said on its own could come back in another alphabet. Now a take of a word or two is heard twice, and tabi keeps the one in your letters.
What’s next
There’s more in Voxtral 3 that tabi doesn’t use yet, and Mistral keeps shipping. I’ll keep building on both.
All of this is in the latest tabi, free to download for the Mac.
Everything tabi does with your voice is routed to Mistral, and to nobody else. tabi is my own project. Mistral doesn’t build, own or endorse it.