Windows
Windows 11
Download for WindowsFor forty years we explained ourselves to computers in symbols. Say it in words.
Voice input, an agent and autocomplete for Windows. On top of any program.
Works on top of what you already use
CursorClaudeChatGPTCopilotVS CodeJetBrainsZedSlackTelegramWhatsAppGmailNotionObsidianDocsFigmaLinearChromeCursorClaudeChatGPTCopilotVS CodeJetBrainsZedSlackTelegramWhatsAppGmailNotionObsidianDocsFigmaLinearChrome
Speed
Talking is faster than typing. Noticeably faster
Speech runs at 150 words a minute. Typing, at 40.
By voice150 wpm
By keyboard40 wpm
The left-hand clock runs on everything: the speech, and then recognition, formatting and insertion - the same three phases the first screen shows. The program shows no draft while you speak either. On the right, typing at forty words a minute with one typo.
The second step
Anyone can transcribe. The difference is what happens next
It sees where you corrected yourself and where the thought ended.
As heard
so um I looked at that report uh on sales for the third no for the fourth quarter .
“The third - no, the fourth” is a self-correction. Only what you meant stays in the text.
Talking aloud
Not commands - a conversation
It hears you while you speak and answers aloud. Interrupt it.
YouOpen Spotify and put on something calm
AssistantOpening Spotify
YouAnd remind me about the call in half an hour
AssistantGot it. I will tell you at seven, while this conversation is open
1.4 seconds from the end of your sentence to the first sound back
An assistant with hands
Saying it is doing it
You say what you need. Up to three workers get on it.
Midri VoiceAssistant permissions
What the assistant is allowed to do on this computer
This is your computer and your files. Turn on only what you intend to use: a permission can be taken back at any time, but what has been done will not undo itself.
How often to askPermissions decide what is allowed. This decides how often the assistant stops to ask
Always askWork alone, stop at the irreversibleDo not ask
See the screenA screenshot and window contents. Without it the assistant answers blind
Launch programsOpen and switch between programs that are already installed
Work with filesRead, create, change and rename your files. Deletion goes to the recycle bin, not past it
Run commandsPowerShell commands. Destructive ones are blocked inside the program and will not run under any permission
Use the browserOpen pages, read them and click - in the worker's own browser. It does not see your tabs or your sign-ins: it works only where no sign-in is needed
Yours
It looks the way you want it to
Four shapes, seven colour sets. Pick one, or watch it try them on.
Tap a shape or a colour
When you type anyway
Half the sentence was already in your head
It finishes your sentence in your window. Agree with Tab.
Slack
Acme
Channels
#general
#design
#releases
#random
#team
A
Ann10:12
Pushed the fix for the export job, should be quicker now.
M
Max10:15
Nice. I will re-run the nightly and check the numbers.
Re-ran the nightly after your fix, the export took four minutes instead of eleven.
1You stop typing
2It offers the rest
3Tab takes it
Code and prompts
Dictate a prompt into your editor - terms stay terms
“Use state” becomes useState. Names come out as code.
Assistant
~/projectproject
src
components
TaskList.tsx
Sidebar.tsx
hooks
useTasks.ts
package.json
1import { useTasks } from "../hooks"
2
3export function TaskList() {
4 const { tasks, done } = useTasks()
5
6 // one card per task
7 return tasks.map((task) => (
8 <Card key={task.id} done={done(task)} />
9 ))
10}
AssistantReady. Describe the task.
Claude Sonnet
Code names restored from pronunciation and from what is open on screen.
mainTypeScriptLn 7, Col 22
Works on top of what you already useCursorClaudeChatGPTCopilotVS CodeJetBrainsZed
Translation as you speak
Speak your language - it lands in theirs
Translation is not a step. It is the same key press.
Slack
Acme
Channels
#general
#design
#releases
#random
#team
D
Dana13:48
Pushed the fix for the export job, should be quicker now.
I
Ivan13:55
Nice. I will re-run the nightly and check the numbers.
M
Mark14:02
Could you take another look at the pull request?
ENHi! Thanks for the edits, I've fixed everything and sent it again.
Insert in: ES
Message #team
Someone else's text
Select it - and read it in your own language
Select someone else’s text and the bar appears above it.
Telegram
Search
J
James19:12
The rollout is blocked until we sort out the migration - I'd rather ship late than break prod.
D
Design18:20
sent the mockups
A
Anna17:04
ok, tomorrow then
R
Releases16:41
build 412 is up
J
James
last seen recently
Today
Did the staging deploy go through?
18:54
Yes, went out an hour ago - logs are clean.
18:57
Good. One thing though.
19:03
The rollout is blocked until we sort out the migration - I'd rather ship late than break prod.
19:12
Message
Written chat
Sometimes you cannot speak out loud. You can still ask.
A hotkey over any window. Same memory as the voice.
Quarterly report.docx
You are working. No browser tab open, nothing to switch to.
Memory
Not a list of facts, but a web
Not rows in a file but a web of links: neighbours come too.
Midri VoiceConversationsMemory
Search the memory
9 nodes·Kept on this computerShow all
150wpm
Conversational speaking pace
1.4s
From your last word to a spoken answer
25languages
Translation languages on insert
What it can do
One assistant instead of ten windows
Twenty-three things, in three groups, in one program.
The assistant
What it does when you speak to it.
10Answers in a voice, not in text
It hears you while you are still speaking and answers in its own voice, with no transcription in between. About a second and a half to the first sound back.
“Open Spotify” - and it is open
It sees everything installed and launches by name. Ask for one already running and it brings that window forward instead of opening a second copy.
Its own browser, not your tabs
The assistant has a separate Chromium with its own profile: it reads pages, clicks and fills fields only in there. You sign in to the sites you need once, inside it. The browser is downloaded separately: 149 MB to fetch, 344 MB on disk.
Thirty voices to pick from
Low and calm, lively and quick, soft and unhurried - the voice changes from the orb menu and from settings. Two of the thirty work on the free plan.
Sees what you see
Share your screen and ask about what is on it: read an error, walk you through steps, make sense of an unfamiliar window.
Reminds you aloud, in conversation
“Remind me in half an hour” - and in half an hour it says so out loud, as long as the conversation is open. Not a notification, just a sentence in the middle of the talk.
Dictation
What happens between your voice and the text.
9Works with no internet
Recognition runs on your computer. Lose the connection and dictation keeps going, cleaned by rules. Your speech is never sent anywhere.
Your names and terms
Forty-five well-known names such as GitHub, macOS and Kubernetes are restored automatically. Add the rest as a list with no length limit - matches are found by sound.
Your screen is a hint
Names from the active window join the dictionary for the current phrase. “Slack” becomes Slack exactly where you are talking about it.
Any key or combination
Right Alt, Caps Lock, Ctrl+Shift - anything, a single key included. Hold it while you talk, like a walkie-talkie.
Your own writing rules
“Business tone”, “no exclamation marks”, “short sentences” - rules are written in words and applied to every phrase.
Nothing gets lost
Everything you dictated stays at hand and can be pasted again in one click. Stored only on your computer.
On your machine
How it sits there when you are not using it.
4Nine sections, not one little box
Dictation, text processing, chat, Live AI, connectors, history, appearance, system, and the subscription tile. The assistant's permissions and its workers live inside “Live AI”, where you switch the conversation on. Theme, twelve accent colours, window material, and the edge the panel sits on.
A thin strip when idle
Until you touch it the panel is a strip at the edge of the screen. Hover and it unfolds into the microphone, conversation, assistant and settings - then tucks away again.
You can bring your own key
If you would rather not pay through us, paste your own key from OpenAI, Anthropic, Google or another provider and its models line up beside ours for the workers. The key sits in the Windows Credential Manager and never travels to us.
Works off to the side
The panel does not take focus, does not pop over your work, and does not move the cursor. The assistant opens a window of its own in one case only - when you have granted it the browser and it is using it.
Install
Your speech never leaves your computer
Recognition runs on the local model - no internet, nothing sent anywhere. An account is needed to use the program, and it opens what runs on our side: smart processing, translation, talking aloud, and the assistant.
macOS
Planned
Not ready yetLinux
Planned
Not ready yetThe program is in a closed beta; there is no public build yet. Create an account and we will write to it as soon as the build can be handed over. macOS and Linux versions are planned - work on them has not started.
Pricing
Your computer computes it free. Our server computes it for money
Recognition runs on your machine and costs us nothing. We charge for what runs on our side: processing, translation, talking aloud and the agent.
Yearly is 20% cheaper
Plus
I write all day
$5.99per month
Choose- Dictation
- unlimited
- Autocomplete
- unlimited
- Agent
- 1000k
- Machines
- 3
- People
- 1
Pro
I write and I delegate
$12.99per month
Choose- Dictation
- unlimited
- Autocomplete
- unlimited
- Agent
- unlimited
- Machines
- 5
- People
- 1
Ultra
There are several of us
$39.99per month
Choose- Dictation
- unlimited
- Autocomplete
- unlimited
- Agent
- unlimited
- Machines
- 10
- People
- 5
Allowances reset on the first of the month, together with your billing.
Questions
The things people ask before downloading
Midri Voice is a voice assistant for Windows. You hold a key and speak - the text arrives finished in whatever window you were typing in. Or you talk to it out loud and it answers in a voice, opens apps, and sorts out your files.
It is free while in preview. There are no payments on the site yet and no card is asked for anywhere.
Yes. Sign-in is through Google and takes one click - there is no password to invent or forget. The account is what the paid plans and the usage allowance are attached to.
Dictation is recognised on your own machine and the audio does not go anywhere. A spoken conversation is different: there your voice travels through the service, because otherwise nothing could answer you.
Windows 10 and Windows 11 today. Other systems are coming - the assistant itself is not tied to Windows.
Any window with a text field: Cursor, VS Code, JetBrains, the terminal, Telegram, Slack, Gmail, Notion, a browser. It remembers the field you clicked before you started speaking, so the text lands there even if the window moved.
Built-in dictation writes down sounds. This one understands what you meant: it punctuates, drops the ums, keeps code and names intact, translates as it goes, and remembers what you told it yesterday.
Try it - it's quicker than finishing this page
Your speech is recognised on your own computer. The account is for what is computed on our side: smart processing, translation, talking aloud, and the assistant.
Windows 11 · free while in preview