DictateStack — dictate, snippets and text polish into any field

Public Beta — actively under development. Features and API may change at short notice. Feedback and bug reports welcome.

What is this?

Many dictation features from the big vendors send spoken audio to US servers for transcription. For notes about patients, clients or internal processes that isn’t an option — and in everyday life there’s little reason to ship every sentence you speak into someone else’s data centre.

DictateStack is a small, focused voice-to-text app for Windows. The Android build is in progress, iOS is planned. Transcription runs through our own backend on a Hetzner server in Germany — the model inference itself with Mistral in France. No US path, no telemetry, no app store. Updates ship directly from us.

Three things that make DictateStack

System-wide dictation A push-to-talk hotkey running in the background — hold the key, speak, the text lands in the focused field. Including third-party apps (mail, messengers, notes, browser forms). Currently for Windows; Android and iOS to follow.

Auto-transforms Seven presets, switchable on the fly: Light Clean-Up (filler words, capitalisation), Medium Polish, High Polish, Mail-Style (salutation, closing), code (transcript cleaner for code blocks), Shorten 50 % and Bullet Points. Auto-apply runs after each dictation when active.

Voice snippets (Shorts) Say snippet greeting and Best regards, Steve lands in the field. Placeholders like {{1}}, {{2}} allow parameterised snippets (e.g. snippet salutation → Hi {{1}},).

A look at the desktop app

Tech

• Backend — Rust + Axum, lean REST API on our own Hetzner server (DE), reachable at dictate.stackschmiede.de
• Transcription — Mistral voxtral-small-latest via Mistral’s EU API (FR)
• Transforms & polish — Mistral Small / Large; per-call cost visible in the HUD
• Auth — JWT (15 min) + refresh token (7 d), SHA256-hashed tokens in SQLite
• Privacy — audio only in RAM (never on disk), Privacy Mode skips the DB insert entirely, retention 24 h / never / normal controllable via header
• Windows client — Tauri 2 with global hotkey, push-to-talk, auto-paste into focused field; signed direct download — no Microsoft Store
• Android — Flutter build in progress (Accessibility Service for system-wide dictation without keyboard switching)
• iOS — planned (Custom Keyboard, later a TestFlight track)
• Language switchable, default DE/EN; more languages prepared

Status

Public Beta — actively under development. Backend instance running at dictate.stackschmiede.de, Tauri Windows client available as direct download (code signing in preparation). Android client in progress, iOS variant planned. There is intentionally no app-store distribution — updates ship directly from us.

Next up: finalising the custom lexicon (domain vocabulary), hardening the Android Accessibility Service, prototyping the iOS Custom Keyboard, and an in-browser demo on this page (see below).

Positioning

DictateStack is intentionally small. No assistant, no AI chit-chat, no US cloud paths. A tool for people who want to type text faster — and keep control over their speech-to-text data.

Distribution without an app store: the Tauri Windows build ships directly from us, no platform gatekeepers, no 30 % cut to Apple or Google. To be honest about the downside: no automatic discovery in a store, no one-click install, and code-signing trust takes time to build.