Whisperpad
Press a key, speak, and the words appear wherever your cursor already is. The speech model runs on your own machine. No audio ever leaves your PC — there is no server to send it to.
Available now Download for Windows, 714 MB
What it does
-
Everything on device
Transcription happens on your own hardware. Your microphone audio is never uploaded, never stored, and never seen by anyone, including me.
-
Works in every app
One global hotkey, anywhere in Windows. The text lands in whatever field has focus — a browser, an editor, an email, a chat box.
-
Offline once set up
Download a model once and the app never needs the internet again. It keeps working on a machine with no connection at all.
-
Pick your model
Seven to choose from, from a 140 MB one that runs on any CPU up to full large‑v3. Trade accuracy against speed on the hardware you actually have.
-
Ninety‑nine languages
The multilingual models handle speech in a great many languages, or you can pin it to one for better accuracy.
-
Built for a screen reader
Raw Win32 controls throughout, so NVDA reads the window, the settings and the model manager properly rather than guessing.
How it works
- Put your cursor wherever you want the text.
- Hold the hotkey —
Ctrl+Shift+Dunless you change it — and speak. - Release. The audio is transcribed on your machine and pasted straight in.
Choosing a model
Models download once, straight from Hugging Face, and are then yours on disk. Bigger is more accurate and slower; a graphics card makes the large ones practical.
| Model | Size | Best for |
|---|---|---|
| base | ~140 MB | Any CPU, older machines. Fast, decent accuracy. |
| small | ~460 MB | A good balance on modest hardware. |
| distil‑large‑v3 | ~1.6 GB | The default. Best speed against accuracy without a GPU. |
| distil‑large‑v2 | ~1.6 GB | English only, very fast, excellent for English speech. |
| medium | ~1.5 GB | Noticeably better than small. Wants a GPU for real‑time use. |
| large‑v3‑turbo | ~1.6 GB | Near large‑v3 quality at roughly twice the speed. |
| large‑v3 | ~3.1 GB | The most accurate. Needs a GPU with 4 GB or more of memory. |
What you need
- Windows, and a microphone.
- Enough disk space for the model you pick.
- An internet connection once, to fetch that model. After that, never again.
- A graphics card is optional, and only matters for the larger models.
Getting it
714 MB, hosted on GitHub. Nothing to install, and nothing to sign up for. All releases and release notes.
- Extract the zip anywhere you like.
- Open the
Whisperpadfolder and runWhisperpad.exe, keeping the_internalfolder beside it. - Windows will warn about an unsigned program. Choose More info, then Run anyway. That warning means the software has not been through paid code signing, not that anything is wrong with it.
- Open Models and download one — base to try it quickly, or distil‑large‑v3 for everyday use.
- Put your cursor where you want the text, hold
Ctrl+Shift+D, and speak.
Why the download is large
Most of it is the NVIDIA CUDA runtime, so the app can use your graphics card — that is what makes the bigger, more accurate models fast enough to use as you speak. The rest is the Python runtime and the audio libraries.
Everything needed to run is in the zip and nothing else is: no installer, no Python to set up, no separate runtime to hunt down, and no source code — just the compiled program and the libraries it loads.
It works without an NVIDIA card as well, using your processor instead. On those machines stick to base or small. The speech models themselves are downloaded separately from inside the app, so you only fetch the one you actually want.