Source Download

Typed, without
the typing.

Talk normally. Finished text lands at your cursor, in any app. Free, open source, and it runs on your own computer.

Example: you say “um, I'll send the, uh, draft on, like, Tuesday, no wait, Wednesday.” Typd types “I'll send the draft on Wednesday.”

$0No account, no subscription, no word limit.
0.5 sFrom key-up to text on short dictations, on my 4 GB laptop.
Wi-Fi offIt keeps working. Nothing is shared unless you say yes.
1 modelIt hears and writes in one step. No second AI to tidy up.

One model, not two.

Most dictation apps run one model to hear you and a large language model to clean up the transcript. I wondered why it couldn’t be one. Typd is a single model that hears you and writes the finished text in one step. I built it on Qwen3-ASR 1.7B and trained it to write what you meant, and it runs on your own computer.

Model
Built on Qwen3-ASR 1.7B and trained for dictation. 8-bit, 2.2 GB, downloaded once.
Runs on
A Windows PC with an NVIDIA GPU of 4 GB or more (it uses about 2.5 GB), or an Apple silicon Mac.
Network
Dictation works offline. Typd goes online for the model and for updates. Nothing is shared unless you say yes when Typd asks. Privacy
Code
The app is open source, GPL-3.0, on GitHub. The model has its own licence.
Cost
Free. No account, no subscription.

It writes what you meant.

Change your mindLet’s meet at four, no wait, five uh in the lobby

Let’s meet at 5 in the lobby.

Think out loudSo um can you like send me the the numbers before Friday

Can you send me the numbers before Friday?

Make a listTo-do list bullet point making my bed bullet point walk the dog bullet point finish homework

To-do list:

  • Making my bed
  • Walk the dog
  • Finish homework

Write an emailHi Priya um thanks for sending this over I’ll read it tomorrow morning best Mehmet

Hi Priya,

Thanks for sending this over. I’ll read it tomorrow morning.

Best,
Mehmet

Every app.

Slack, Gmail, VS Code, Notepad. If there’s a cursor, it works.

Your words.

It reads names off your screen and learns from your corrections.

One tap.

Tap a key, talk, tap again. That’s the whole app.

It’s still early.

Mostly my voice.

I’ve tested it mostly on my own dictations. Yours may go differently.

Still learning.

Long lists and bigger changes of mind trip it up, and it gets some names wrong.

Picky hardware.

It needs an Apple silicon Mac, or a Windows PC with an NVIDIA GPU of 4 GB or more.

I’d really like to hear where it gets things wrong: email me or open an issue on GitHub. Where it stands →

— Mehmet

Download

Free. Offline.
Open source.

FreeWorks offlineWindows 10/11 with an NVIDIA GPUmacOS 13.3+ on Apple silicon

Install steps and requirements →

The Windows app isn’t signed with a paid certificate yet, so Windows asks you to confirm the first time you open it. The Mac app is signed and notarized. Checksums are on the install page, and you can build it from source.