Windows 10 & 11 · Free & open source

Make Waves.
Not Keystrokes.

Private voice dictation that types wherever you are. Press Alt + W in any app, talk, and clean, punctuated text lands right where your cursor is. Speech recognition and AI run on your own PC, so your voice never leaves it.

Download for Windows View on GitHub

MIT licensed · 64-bit Windows · 8 GB RAM recommended · graphics card optional

WaveScribe main window, light theme WaveScribe main window, dark theme
// how it works

Three steps. Any app.

Email, Word, Slack, a browser or your code editor. If it has a text cursor, WaveScribe can type into it.

01

Press Alt + W

Click into any text field and hit the shortcut. Prefer to hold a key while you speak? Turn on push to talk.

02

Just talk

A small indicator appears beside your text cursor and moves with your voice. Speech is transcribed phrase by phrase as you go.

03

Press it again

Your words appear right where you were typing: punctuated, capitalized, with the ums and uhs removed.

The WaveScribe indicator next to the text cursor while you talk, and while it transcribes
// why wavescribe

Built to stay out of your way.

Dictation that is fast enough to trust and private enough to use for anything.

Private by design

Speech recognition and every AI feature run locally. Your voice and your text never leave your PC.

Fast

Phrases are decoded while you are still talking, so the text appears moments after you stop.

Smart edit

Drops self-corrections ("Tuesday, no wait, Wednesday") and false starts, turns spoken lists into lists, and understands "new paragraph" and "scratch that".

Custom words

Teach it the names and jargon it mishears, so "gooey" becomes "GUI" every time.

Mini mode

Shrink WaveScribe to a small pill that floats above your work and shows when it is listening.

Updates itself

New versions download in the background. One click on "Update" installs it, and WaveScribe reopens on its own.

// voice commands

Select. Speak. Done.

Highlight text in any app, press Alt + Shift + W and say what to change. The selection is replaced with the result. With nothing selected, it writes what you ask for.

Selectedhey so the launch is moving to friday because qa found a couple bugs, can everyone update there calendars thanks
Alt + Shift + W
Make this more formal
ResultHi team, the launch has moved to Friday so QA can resolve a few remaining bugs. Please update your calendars accordingly. Thank you.
// ai chat

A private AI, offline.

Chat with open-source models like Llama, Qwen, Gemma and Phi, running entirely on your PC.

  • Formatted replies with headings, lists and code.
  • Dictate your questions instead of typing them.
  • No account, no API key, and nothing sent to the cloud.
AI chat with a formatted reply from a local model
// models that fit your pc

It knows what your hardware can run.

WaveScribe checks your PC and the memory that is free right now, then recommends models that will run well on it.

  • Three independent choices: voice to text, text cleanup and AI chat.
  • One-click presets: Fast, Balanced or Precise.
  • Clear fit signals: whether each model runs well now and fits on your graphics card, with the best one marked Recommended.
Settings with the three model choices and presets Model picker showing recommendations based on free memory

What goes online? Almost nothing.

Your audio and text stay on your PC. WaveScribe connects to the internet only to download the models you choose (from Hugging Face), to check GitHub for updates, and to load its font.

// keyboard shortcuts

Hands on the keys, barely.

Alt + WStart or stop dictating (change it in Settings)
Hold the shortcutPush to talk, when enabled in Settings
Alt + Shift + WVoice command on the selected text
Enter / Shift + EnterSend a chat message / new line
// under the hood

How dictation actually works.

A streaming pipeline built so the wait after you stop talking is as short as possible.

Capture16 kHz mono audio is split into phrases at natural pauses. Long run-on speech is cut at the quietest point, never mid-word.
Transcribe liveEach phrase is decoded while recording continues, with Moonshine or Whisper on a multi-threaded CPU or WebGPU.
StitchPhrases are joined back together, removing false full stops the model adds at forced cuts.
Clean & pasteFiller removal, custom words, capitalization and optional local-LLM cleanup, then a native paste at your cursor.
ElectronTypeScriptVitetransformers.jsMoonshineWhisperWebGPUnode-llama-cppWin32 SendInputelectron-updater
WaveScribe mini mode while recording

Stop typing. Start talking.

Free, open source and private. Download the installer, press Alt + W, and make waves.

Download WaveScribe All projects