Voice to Text

A local, offline voice-to-text (dictation) app for Windows 11. It lives in the system tray; a global hotkey toggles dictation, audio from the selected microphone is transcribed locally with Whisper, and the text is typed into whatever window has focus. Everything runs on-device — no cloud, no API keys, and your audio never leaves your PC.

Install

Download the latest VoiceToText-Setup-x.y.z.exe from Releases and run it (per-user install, no admin needed). A first-run wizard picks your microphone and hotkey; the speech model (~1.6 GB) downloads once and everything is offline from then on. Updates are offered automatically from this repository's releases when enabled in Settings.

Stack

Concern	Choice
Language / shell	C# / .NET 10, WinForms tray host
Speech-to-text	Whisper.net (whisper.cpp), model large-v3-turbo
GPU acceleration	Vulkan runtime (works on AMD/Intel/NVIDIA; no CUDA needed)
Audio capture	NAudio (WASAPI) → 16 kHz mono
Global hotkey	Win32 RegisterHotKey (toggle)
Text injection	Clipboard paste (Ctrl+V via SendInput)

All core dependencies are MIT-licensed.

Requirements

• Windows 11
• .NET 10 SDK
• A Vulkan-capable GPU (optional — falls back to CPU)

Build & run

dotnet build
dotnet run --project src/VoiceToText

On first launch the large-v3-turbo model (~1.6 GB) is downloaded once to %APPDATA%\VoiceToText\models. After that it loads from disk.

Usage

• The tray icon is blue (idle), red (recording), amber (transcribing).
• Press your hotkey to dictate — press-to-toggle or hold-to-talk (pick in Settings). Press-to-toggle can auto-stop after a pause; the transcription is pasted at your cursor.
• Double-click the tray icon for the dashboard: usage stats, Settings (microphone, speech model, hotkey, activation), Text rules (find→replace corrections + spoken commands like "new line"), opt-in dictation History, and About/diagnostics.
• The hotkey can be a single key (e.g. an extra/macro key or F13) or a modifier combo. Esc/Tab/Enter stay reserved for the dialog.
• Tray menu → Check for updates… checks the configured update source now.

Verify the engine (no mic needed)

dotnet run --project src/VoiceToText -- --selftest path\to\16k-mono.wav selftest-output.txt

Writes the transcript, timing, and which native runtime loaded (e.g. Vulkan).

Build a standalone .exe (no .NET install needed)

.\publish.ps1

This produces a self-contained app in publish\:

publish\
├─ VoiceToText.exe      # ~50 MB — all managed code + the .NET runtime bundled
└─ runtimes\            # ~48 MB — Whisper native libs (Vulkan + CPU)

Double-click VoiceToText.exe on any Windows 11 x64 machine — no .NET, no DOTNET_ROOT, nothing to install. Keep the runtimes\ folder next to the exe (Whisper loads its native libraries from there). To move/share the app, copy the whole publish\ folder (or zip it).

The speech model is still downloaded on first run to %APPDATA%\VoiceToText\models (it is intentionally not bundled into the exe).

Build the installer

A per-user installer (no admin needed; Start-menu shortcut, optional desktop shortcut and start-on-login, clean uninstall) is defined in installer\VoiceToText.iss (Inno Setup).

.\publish.ps1                       # 1. produce publish\
iscc installer\VoiceToText.iss      # 2. compile -> installer\Output\VoiceToText-Setup.exe

If you don't have Inno Setup, its compiler ships in the Tools.InnoSetup NuGet package (no install required) — see the ISCC.exe inside that package. The installer version is read automatically from the published exe, so just bump <Version> in the csproj.

Auto-update

The app updates itself from a configurable update source — by default this repository's GitHub Releases (https://github.com/lukeeexd/voice-to-text/releases/latest/download), but a local/UNC folder works too (handy for development). Enable it in Settings (off by default). On startup and via tray → Check for updates…, the app reads latest.json from the source; if it names a newer version it offers to install it, downloads/copies the setup to a local staging dir (verifying its SHA-256), and a small relauncher runs the installer and reopens the app. Your settings and model in %APPDATA%\VoiceToText are untouched.

⚠️ The app runs the installer the source provides, so only use a source you control and trust. The installer isn't code-signed yet (see roadmap), so the feature is off by default and shows a one-time consent prompt when enabled.

Publishing a release

# bump <Version> in src\VoiceToText\VoiceToText.csproj, then:
.\publish.ps1
iscc installer\VoiceToText.iss
# rename the setup with its version and compute its hash, write latest.json, then:
gh release create v<ver> VoiceToText-Setup-<ver>.exe latest.json --title "v<ver>" --notes "What changed"

latest.json (uploaded as a release asset next to the setup):

{
  "Version": "0.9.0",
  "SetupFileName": "VoiceToText-Setup-0.9.0.exe",
  "Sha256": "<lowercase sha256 of the setup exe>",
  "ReleaseNotes": "What changed"
}

(SetupFileName must be a plain file name living next to the manifest — the updater rejects paths.)

Known limitations

• Elevated windows: Windows UIPI blocks both the hotkey and paste while an admin-elevated window is focused. Dictation works in normal apps (Word, Outlook, browsers) but not into admin-elevated windows.
• Dictation uses a single transcription pass on stop (push-to-talk / toggle), not live streaming.

Roadmap

• Upgrade auto-stop to Silero VAD (more robust against background noise than the current energy-based detector)
• Live/streaming partial results
• Code-sign the installer + verify Authenticode/publisher in the updater (so auto-update trusts the binary, not just a manifest hash)

License

MIT