Local Windows voice-to-text dictation app. Hold a global hotkey, speak, release — text is transcribed on-device and inserted into the focused application.
This repository is a Flutter + native C++ Windows port. The original FluidVoice macOS app (Swift) remains in-tree as reference only and is not what you build or run on Windows.
| Layer | Tech |
|---|---|
| UI / app shell | Flutter (Windows desktop) under flutter_app/ |
| Mic, hotkeys, STT, inject, tray, overlay | Native plugins under native_plugins/ |
| Speech | whisper.cpp (ggml) and Parakeet TDT (ONNX via sherpa-onnx) behind one speech_runtime facade |
- Push-to-talk global hotkey (default F8, configurable in Settings)
- WASAPI microphone capture with device picker
- Whisper
tiny.en/base.en(ggml) - Parakeet TDT
parakeet-tdt-0.6b-v2-int8(English ONNX, offline batch) - Text injection into other apps (SendInput → clipboard → UIA)
- Live overlay, floating dictation pill (optional), system tray (close quits; minimize-to-tray is optional)
- Settings & transcript history (AppData JSON)
- Launch at login, always-on-top, optional acrylic
- Optional cloud AI output modes (raw / enhance / rewrite / write) via API key in Windows Credential Manager
- Optional DeepL translation — bring your own DeepL API key; transcribe locally, translate, inject the translation
- Optional local loopback API on
127.0.0.1:47733 - Portable zip and Inno Setup installer
This is not a full port of macOS FluidVoice. See docs/WINDOWS_RELEASE_NOTES.md for a clear ported vs not ported list to use on GitHub Releases with the installer / portable zip.
- Flutter stable with Windows desktop enabled (
flutter config --enable-windows-desktop). - Visual Studio 2022 with the “Desktop development with C++” workload.
- Workspace path without
;. If the real path has a semicolon, use a junction:
cmd /c mklink /J C:\dev\FluidVoice_Port_to_Windows "C:\path\to\real\repo"- Fonts under
flutter_app/assets/fonts/(see Fonts below) — required beforeflutter build/flutter run. - For the Setup installer only: Inno Setup 6.
Prefer Release for usable STT speed:
cd C:\dev\FluidVoice_Port_to_Windows\flutter_app
flutter pub get
flutter run -d windows --releaseHold the hotkey to dictate; release to transcribe and inject. On first use, the selected model downloads (network once). Parakeet is ~400MB.
On Home, the status line shows the active model, e.g. model:parakeet-tdt-0.6b-v2-int8 · backend:Parakeet ONNX.
From the repo root (junction path recommended):
cd C:\dev\FluidVoice_Port_to_Windows
.\scripts\windows\build_portable.ps1
.\scripts\windows\build_installer.ps1| Output | Location |
|---|---|
| Portable folder | dist\FluidVoice-portable\ |
| Portable zip | dist\FluidVoice-portable.zip |
| Setup exe | dist\FluidVoice-Windows-Setup-1.0.0.exe |
Both packages are the same Flutter Release app (optimized binary + DLLs + embedded assets including fonts). Speech models are not bundled; they download into AppData on first use.
Portable (FluidVoice-portable.zip) |
Setup (FluidVoice-Windows-Setup-*.exe) |
|
|---|---|---|
| What it is | Unzip-and-run copy of the Release folder | Inno Setup installer wrapping that same Release folder |
| Install | None — extract anywhere (USB, Desktop, etc.) | Copies into a Program Files–style app folder |
| Shortcuts | None | Start Menu (+ optional desktop icon) |
| Uninstall | Delete the folder | Windows Apps & features / uninstaller |
| Who it’s for | Trying the app without installing, or keeping a self-contained folder | Typical end-user install |
More toolchain detail: docs/BUILDING_WINDOWS.md.
Font .ttf files live under flutter_app/assets/fonts/ and are gitignored (large binaries). They are declared in flutter_app/pubspec.yaml and embedded into Release builds, so downloaders of the portable zip or Setup exe already get them — no separate font download.
| File | Role |
|---|---|
FluidUI-Light.ttf |
UI (weight 300) |
FluidUI-Regular.ttf |
UI (400–600) |
FluidUI-Bold.ttf |
UI (700) |
FluidUI-Italic.ttf |
UI italic |
NotoSans-Regular.ttf / NotoSans-Bold.ttf |
Multilingual Latin/Cyrillic/Greek transcript fallback |
NotoSansArabic-Regular.ttf |
Arabic fallback |
NotoSansDevanagari-Regular.ttf |
Devanagari fallback |
CJK falls back to Windows UI fonts at runtime (no bundled CJK TTFs).
- Place the FluidUI files in
flutter_app/assets/fonts/(same set used for the Windows port branding). - Download Noto Sans / Noto Sans Arabic / Noto Sans Devanagari from Google Fonts (SIL Open Font License) and copy the Regular/Bold TTFs named exactly as in the table above.
- Run
flutter pub getthen build as usual.
If fonts are missing, flutter build windows will fail on the missing asset paths in pubspec.yaml.
Select under Models in the app:
| Id | Backend | Notes |
|---|---|---|
tiny.en |
Whisper (ggml) | Fast English default |
base.en |
Whisper (ggml) | Higher quality English |
parakeet-tdt-0.6b-v2-int8 |
Parakeet via sherpa-onnx | English ONNX int8, larger download |
After changing models, return to Home so the app reloads the selection. Status text shows which backend is loaded.
Turn dictation into a speech → translate → inject workflow. Speech-to-text still runs on-device; only the transcribed text is sent to DeepL when translation is enabled.
- Hold the hotkey and speak (local STT, unchanged).
- Optional AI output mode runs first if not set to
raw(enhance / rewrite / write). - If Translate before inject is on, the text is sent to the DeepL API using your API key.
- The translated text is injected into the focused app (not the raw STT output).
There is no FluidVoice backend for translation — the app calls DeepL directly from the client with a bring-your-own-key setup.
- Create a free or paid API key at deepl.com/pro-api.
- In the app: Settings → Translation.
- Enable Translate before inject.
- Choose a target language (e.g.
DE,FR,EN-US). - Paste your DeepL API key and click Save.
Keys are stored in Windows Credential Manager (same as the optional AI API key), not in plain settings JSON.
| Key type | Endpoint used automatically |
|---|---|
Free (key ends with :fx) |
api-free.deepl.com |
| Pro | api.deepl.com |
If translation fails (missing key, network, quota), the app falls back to the pre-translation text and shows the error in the status detail on Home.
flutter_app/ # Flutter Windows application
native_plugins/ # C++ plugins + shared plugin API headers
scripts/windows/ # Portable + installer + smoke tests
docs/ # Architecture, build, port plan
WINDOWS.md # Short Windows quick start
Sources/ … # Upstream macOS Swift (reference only; not built here)
More detail:
- WINDOWS.md — Windows quick start
- flutter_app/README.md — Flutter app notes
- docs/BUILDING_WINDOWS.md — toolchain and packaging
- docs/ARCHITECTURE.md — system design
- docs/WINDOWS_PORT_PLAN.md — port status
- native_plugins/speech_runtime/onnx_runtime/README.md — Parakeet / ONNX notes
- Windows 10 or 11 (x64)
- Flutter stable + VS 2022 C++ desktop workload (to build)
- Microphone
- ~100MB+ disk for Whisper tiny, or ~400MB+ for Parakeet
- Network on first model download
Dictation STT runs on-device. Audio stays local.
Optional features send text (not raw audio) to third-party APIs only when you enable them and provide your own key:
| Feature | What leaves your PC |
|---|---|
| AI output modes (enhance / rewrite / write) | Transcribed text → your configured OpenAI-compatible endpoint |
| DeepL translation | Transcribed text → DeepL API (api-free.deepl.com or api.deepl.com) |
With translation and AI disabled, transcripts stay on the machine (AppData history JSON).
This project is licensed under the GNU General Public License, Version 3.0 (GPLv3).
The macOS Swift tree retained in this repo is upstream FluidVoice reference material; the Windows product path is flutter_app/ + native_plugins/.