VoiceTyper

A small Windows tray app that turns speech into typed text. Hold down the record key, speak, release it, and the transcript is typed into whatever field currently has focus — no clicking required.

Transcription runs through your choice of provider: Mistral Voxtral, Google Gemini, or both at once for side-by-side comparison.

How it works

1. Hold Right Alt (AltGr) to start recording.
2. Speak.
3. Release the key to stop recording. The audio is sent to the configured transcription provider:
   • Mistral — the Voxtral transcription API (voxtral-mini-latest).
   • Google — the Gemini Interactions API (gemini-3.5-transcribe), with the audio sent inline in the request.
   • Both — both providers in parallel.
4. The returned text is typed into the currently focused field via the keyboard (no mouse involved).

In Both mode the two transcripts are typed one after the other, each under a [Mistral] / [Google] label and joined by the configured separator. If only one provider answers, its transcript is typed anyway and a notification names the one that failed — you never lose a dictation to a single provider outage.

Also holding the summary modifier (Right Shift by default) at any point during the hold upgrades that recording to transcript + summary: release the record key and VoiceTyper sends the transcript to your chosen summary model — a Mistral chat model or a Gemini model — for a summary, then types the transcript, a blank line, and the summary. The modifier is sticky for the hold — press it any time before releasing the record key, or release it early, and the upgrade still applies — but release it before the text types to avoid upper-cased output. It does nothing by itself; the record key must be held for anything to happen at all. If the summary call fails the transcript is still typed and a notification says so — a failed summary never costs you your dictation.

The summary gesture does nothing in Both mode. Comparing two transcripts and summarising are separate jobs; with the provider set to Both the modifier is ignored and you get the two labelled transcripts only.

A tray icon shows the current state (idle, recording, processing, paused, or error) and gives you a menu to pause/resume, open Settings, or quit.

A note on typing speed. Text is typed via synthesized keystrokes, one character at a time with a small delay between each (2 ms). A long dictation can occupy the keyboard for a few seconds, and both the summary command and Both mode roughly double the character count, so they take about twice as long. Avoid typing yourself into the focused field until the tray icon leaves the "processing" state.

Requirements

• Windows only. VoiceTyper relies on Windows-specific APIs (global key listening, keyboard synthesis, DPAPI-backed secret storage, login items) and is not supported on other platforms.

• Node.js 24 or later.

• A microphone.

• An API key for at least one provider:

  • a Mistral API key with access to the audio transcription endpoint, and/or
  • a Google Gemini API key.

  Both mode needs both keys.

• An internet connection — transcription is not local. Every recording is uploaded to the selected provider's API for processing, and no audio is transcribed on-device.

Install

git clone https://github.com/Nkay/voice-typer.git
cd voice-typer
npm install

Run in development

npm start

This launches the Electron app. On first run, no API key is configured yet, so you'll get a notification prompting you to set one — open the tray icon's Settings menu item to add it.

Configuring your API keys

Right-click the tray icon and choose Settings. Paste your Mistral key into the "Mistral API key" field, your Gemini key into the "Google API key" field — whichever you plan to use — and click Save. Each key is encrypted at rest using Windows DPAPI (via Electron's safeStorage) and stored separately from the rest of your configuration; neither is ever written in plain text to disk.

Saving is validated against what you selected: choosing the Google or Both provider, or a google/ summary model, without a Google key on file is rejected with a message rather than silently failing at record time.

Settings

The same Settings window lets you configure:

• Mistral API key — leave the field blank to keep the currently saved key.
• Google API key — likewise; blank keeps the saved key.
• Transcription provider — Mistral, Google, or Both. Defaults to Mistral.
• Language — auto (auto-detect), de (German), or en (English). Applies to whichever provider is in use.
• Record key — the key you hold for a plain transcript. Defaults to Right Alt (AltGr). Selectable keys are Left/Right Alt, Left/Right Ctrl, Left/Right Shift, F1–F24, and Space.
• Summary modifier — the key that, held along with the record key at any point during the hold, upgrades a recording to transcript + summary. Defaults to Right Shift. Must differ from the record key. Set to — none — to disable the summary command entirely.
• Summary model — which model writes the summary. The list is fetched from your accounts, so it reflects the models you actually have access to, and entries are prefixed by provider: mistral-small-latest (Mistral), gemini-3.7-flash (Google). Only providers with a saved key contribute models; if one provider's list fails to load the other is still shown, with a message naming what went wrong.
• Summary prompt — the instructions sent to that model. Put style rules or worked examples here.
• Separator — what goes between the transcript and the summary, and between the two transcripts in Both mode: a blank line (typed as Shift+Enter twice, so it never submits a chat field), an em dash, or two spaces.
• Microphone — which input device to record from (defaults to the system default device).
• Launch on login — whether Windows should start VoiceTyper automatically when you log in. Off by default.

A note on Space and function keys. VoiceTyper watches keys, it does not intercept them. Binding Space means holding it also types spaces into the focused field, and function keys may trigger whatever the focused app uses them for (F1 help, F5 refresh, F12 devtools). The modifier keys are the safe choice.

Configuration file

Non-secret settings are stored as JSON at:

%APPDATA%\VoiceTyper\config.json

That includes transcriptionProvider ("mistral", "google", or "both") and the provider-prefixed summaryModel. The API keys are stored separately (encrypted, as key.enc for Mistral and google-key.enc for Google) in the same directory and are not part of this file.

Building the installer

Run on Windows (electron-builder cannot cross-compile Windows targets from other platforms):

npm run icons && npm run dist

This generates the tray/app icons and produces both an NSIS installer and a portable executable under dist/.

Running tests

npm test

The unit suite (Vitest) mocks all hardware and network dependencies — audio capture, the global key listener, keyboard synthesis, the Mistral and Gemini APIs, and Windows secret storage — so it runs without any real devices or network access.

Notes on privacy and connectivity

Transcription requires an active internet connection. Recorded audio is sent to the provider you select — Mistral's Voxtral API, Google's Gemini API, or, in Both mode, to both of them; VoiceTyper does not transcribe audio locally.

The summary command sends the transcript text to the summary model's API in a second request, so a summarised dictation reaches a provider twice: once as audio, once as text. If you pick a Gemini summary model while transcribing with Mistral (or the reverse), that second request goes to the other provider.

Neither provider retains the audio as a stored file: the recording travels inside the request itself in both cases, rather than being staged through Google's Files API first.

Because the Google path sends the audio inline, it inherits the API's request-size limit — about seven minutes of speech at VoiceTyper's 16 kHz mono capture. A longer recording is rejected before it leaves your machine, with a notification saying so; the Mistral path has no such ceiling.

License

MIT