JVoice

Hold a shortcut, talk, let go. JVoice turns your speech into text and types it wherever your cursor is, in any Windows app.

• Speech to text with ElevenLabs Scribe v2, through your OpenRouter key.
• Optional AI cleanup through any OpenRouter chat model (default DeepSeek V4.1 Flash): removes um and uh, applies "no wait" corrections, adds punctuation. Or pick Rewrite, and the AI rewrites what you said into clear, polished text.
• One key for everything: you only need an OpenRouter key.

Download

Get the installer from the latest release: download JVoice-Setup-<version>.exe and run it. The installer is not code signed, so Windows SmartScreen may warn you. Choose More info, then Run anyway.

JVoice updates itself from this repo's GitHub Releases. It checks at start, every hour, and when your computer wakes, downloads a new version in the background, checks its signature, and restarts into it when you are not dictating.

Use it

1. Start JVoice and paste your OpenRouter key (sk-or-...). Get one at https://openrouter.ai/settings/keys.
2. Click into any text field, hold Ctrl + Win, speak, and let go.
3. The text is pasted where your cursor is. Press Esc while recording to cancel.

JVoice runs in the tray. Closing the window keeps it running; quit from the tray icon.

Settings

• Shortcut: any combo of two or more keys, or a single F13 to F24 key. Hold to talk, or press to start and stop.
• Microphone: pick an input device and test the level.
• Speech to text: Scribe v2 or Scribe v2 Medical, auto language or a fixed one.
• AI cleanup: on or off, Light cleanup or Rewrite, optional extra instructions, and the model (any OpenRouter chat model).
• Output: type into the active app, or copy only. Keep my clipboard puts back what you had copied, including files and images.
• Dictionary: words JVoice should always spell your way, and what the speech engine writes instead.
• History: every dictation, the raw transcript, and Retry for recordings that failed to transcribe.

Speed test

Open Speed test, pick a length (30 seconds up to 1 hour, a custom length up to 3 hours, or no limit), press Start test and talk. JVoice shows your words per minute and characters per minute while you talk, and the final result when the time is up or you press Stop.

• The speech model (Scribe) writes down what you say. JVoice then counts the words and characters itself. No chat AI is used.
• Characters are letters and numbers. Spaces and punctuation are not counted, since you do not say them.
• The rates use the whole test time, pauses included.
• Results are saved on this computer, with your best marked.

Safety

• If you switch windows while JVoice is working, or still hold keys when it is ready, the text is copied instead of pasted, so it never lands in the wrong place.
• The key is encrypted with your Windows account (DPAPI). Settings, history, the dictionary and speed test results live in a SQLite file in %APPDATA%\JVoice.
• Audio is only kept for dictations that failed, so you can retry them. Deleting the history entry deletes the audio.

Develop

npm install
npm start          # run the app
npm test           # unit tests
npm run smoke      # hidden self-check: database, keyboard hook, Win32, pages
npm run dist       # build the Windows installer into dist/

Publish a release

1. Bump version in package.json and commit.
2. Write the changelog for users in release-notes/v<version>.md. Lines starting with - show in the app's update notes.
3. Run npm run release. It runs the tests, builds the installer, signs it with the Ed25519 key in ~/.jvoice/release-signing-key.pem, pushes the tag, and creates the GitHub release with the installer, latest.yml, the blockmap and the .sig file.

Installed copies only install an update whose .sig matches the public key built into app/updater.cjs, so keep the signing key backed up.

Test modes (no window is shown):

• electron . --transcribe file.wav runs the full pipeline on a 16 kHz mono WAV with the saved key and prints the result.
• electron . --speed-test file.wav <length> <stopAfterMs> <dir> runs the Speed test page offscreen with the WAV as the microphone and the saved key, prints the result, screenshots each stage into <dir>, then deletes the saved result. Length is 30, 60, ... 3600, 0 for no limit, or custom2 for 2 minutes. A stopAfterMs above 0 presses Stop after that long.
• electron . --shots <dir> renders every page and overlay state offscreen to PNGs.
• JVOICE_SHOTS_UPDATE=<version> with --shots also presses Check for updates against the real GitHub feed, as a dev copy of that version, and prints what it showed.
• JVOICE_USER_DATA=<dir> uses a separate data folder.

Built with Electron, uiohook-napi for the global shortcut, koffi for Win32 calls, and node:sqlite.