What tabi can do, beyond dictation
A week of how I actually use it, and what happens under each part.
By Nikhil Bhima, founder of tabi
tabi is a voice app for the Mac. Hold fn, talk, and the words land wherever your cursor is. Start with “hey tabi” and it answers, rewrites or replies instead. Select anything and it reads it to you.
People call it a dictation app. Fair, that’s where it starts. This is everything else I use it for, every day.
One week, counted
Most of my writing starts as talking now
tabi counts what it does on my Mac. These are my numbers for the week of 20 September.
Dictation
Writing by talking
Hold fn, talk, let go. What I said comes out as what I would have typed: the ums and restarts gone, punctuation in, a list turned into a list.
Most of the work happens while I’m still talking, so a short sentence lands about 1.3 seconds after I let go, and even a minute-long ramble barely waits longer.
uh yeah thursday no friday works um I’ll bring the build and we can like walk through it together
Yeah, Friday works. I’ll bring the build and we can walk through it together.
hold fn and talk
- Your voice
Voxtral Realtimehears while you talk
Mistral Mediumtidies as it goes- Your cursor
Styles
Each app writes like itself
The app in front picks the style. An email gets its greeting and sign-off on their own lines, Slack gets lighter punctuation, a text stays loose. A style only changes capitals, punctuation and layout, never my words.
MailHi Sam,
Friday works. I’ll bring the build. See you then.
Nikhil
SlackHi Sam, Friday works, I’ll bring the build, see you then
Nikhil
Messageshi sam! friday works, I’ll bring the build, see you then
nikhil
Prompts
Long prompts, said in full
A good prompt for an AI chat needs the whole picture: who it’s for, what I already have, what to leave out. Typing all that is where I used to cut corners. Now I just say it.
It works the same in a chat in the browser or a coding tool in the terminal, where file names, flags and paths come out exactly as I said them.
I’m presenting tabi at a meetup on Saturday, ten minutes, mostly developers. Help me split it into three parts: what it does, how it’s built, and a live demo. Keep the demo last and leave out pricing.
hold fn and say the whole thing
13 languages
Say it in one language, write it in another
tabi hears thirteen languages and picks up the one you switch into. You can also tell it which language to write in, so you can say it in Hindi and send it in English.
Let’s start tomorrow’s meeting a little later, around eleven.
“कल की मीटिंग थोड़ा लेट, ग्यारह बजे के आसपास शुरू करते हैं”
- Hindi
- Voxtralhears it
- Mistral Mediumwrites it in English
- English
Hey tabi
Asking, rewriting, replying
Start a take with “hey tabi” and it becomes a request. What happens next depends on what’s on screen: nothing selected, and it answers; text selected, and it rewrites it; ask for a reply, and it writes one to the message in front of me.
“hey tabi, how do I say see you Thursday in French”
À jeudi !
Ask
The answer lands at my cursor, or in a small box above the orb when there’s nowhere to type.
so basically the thing is that the release is going to be a bit late because the PDF part needs more testing
“hey tabi, make it shorter”
Rewrite
“The release is late: the PDF part needs more testing.” It replaces what I selected.
are we still on for the demo friday?
Yes, still on for Friday. Could you send me the room link?
“hey tabi, say yes, ask for the link”
Reply
I say roughly what I mean, and the reply is written in the box.
- “hey tabi…”
- Voxtral Realtimehears the request
- Mistral Mediumdoes what you asked
- Your cursor
Read aloud
Hearing instead of reading
Select anything and press ⌃⌘R. A real voice starts reading in about half a second, in a small player I can pause, skip or speed up.
I use it to check what I wrote. My ears catch the clunky sentence my eyes skip. It can even read in a copy of my own voice.
Thanks for the notes on the deck. I moved the pricing slide to the end and cut the second demo, so the whole thing now runs just under ten minutes.
- Selected text
- Voxtral TTSspeaks it as it streams
- The player
Listen
PDFs, books and photos, read to you
Listen is a page in tabi for longer things. I open a PDF, a Word file, an EPUB or a photo of a page, and it reads it with each word marked as it’s said. The page follows the voice.
It can read the whole thing in any of nine languages, or save it as audio for a walk.
Chapter 3. The harbour
The harbour was quiet that morning. Fishing boats sat low in the water, their nets already folded. Nobody on the quay seemed to be in any hurry, and the gulls had the steps to themselves. By noon the tide would turn, and the whole town would move with it.
14- A PDF or photo
Mistral OCRonly for scans- Mistral Mediumto translate
- Voxtral TTSreads it aloud
How it fits together
Four roads, all through Mistral
Everything above is one of four roads, and every stop on them is a Mistral model.
- Voice
- Voxtral
- Medium
- Cursor
- Voice
- Voxtral
- Medium
- Answer
- Text
- Voxtral TTS
- Voice
- File
- OCR
- Medium
- Voxtral TTS
Why this is possible
One lab that can hear, write, speak and read
tabi can do this much because Mistral already had every piece it needed. Voxtral hears and speaks, Medium writes, OCR reads pages. When I wanted tabi to read scanned PDFs, OCR was on the same account, behind the same key, under the same privacy policy. No new company to add, no new place for your words to go.
Your voice makes one trip, to one company, and the recording is thrown away once it becomes text. The longer story is in why tabi runs on Mistral.
Questions
Is tabi only a dictation app?
No. Dictation is where it starts: hold fn, talk, and the words land at your cursor in any Mac app. The same key also answers questions, rewrites selected text and writes replies when you start with “hey tabi”, and tabi reads text, PDFs, books and photos aloud.
What can I do with “hey tabi”?
Start a take with “hey tabi” and it becomes a request. With nothing selected it answers at your cursor. With text selected it rewrites that text the way you asked. Ask it to reply and it writes a reply to the message you selected or the conversation in front of you.
Can tabi read text aloud on a Mac?
Yes. Select text in any app and press ⌃⌘R, and a real voice starts reading in about half a second. For longer things, Listen reads PDFs, Word files, EPUBs and photos of pages, marking each word as it is said, in nine languages.
Which apps does tabi work in?
Any Mac app where you can type: mail, chat, notes, documents, the browser and the terminal. Each app can have its own writing style, so an email reads like an email and a Slack message stays light.
What AI does tabi run on?
Mistral AI, and nothing else. Voxtral hears you and reads aloud, Mistral Medium writes, and Mistral OCR reads scanned pages. Mistral does not build, own or endorse tabi.
More on the questions page, or see what tabi does.