Push-to-talk dictation with live or batch transcription, automatic paste, and an open turn for composing multiple fragments.
Builds produce Undertone-Setup-<version>-x64.exe in electron\release.
Windows SmartScreen may warn about an unsigned build.
- Settings:
%APPDATA%\Undertone - Local engines and models:
%LOCALAPPDATA%\Undertone
On first launch, open Settings → Speech & AI, choose a cloud or local transcription provider, and save or install what it needs. Provider keys stay on this PC and are DPAPI-encrypted on disk.
Hold the dictate shortcut and release it to transcribe, clean up, and paste. Tap it once for hands-free recording, then tap it again to finish and paste. Press Left Alt while recording to stop without pasting and append the result to the open turn instead.
| Action | Default shortcut |
|---|---|
| Dictate (hold or tap) | Left Ctrl+Left Windows |
| Keep recording in open turn | Left Alt while recording |
| Commit open turn | Left Ctrl+Left Alt |
| Scratch last fragment | Left Ctrl+Left Alt+Backspace |
| Discard open turn | Ctrl+Alt+Shift+Backspace |
| Re-paste last commit | Left Alt+V |
| Cancel recording | Esc |
A held release or second toggle tap captures the current paste target. If focus changes before injection, Undertone keeps the turn open for manual commit. Use the tray menu for Settings, pause, or quit.
- General — shortcuts, mic, language, sounds, updates
- Speech & AI — STT and cleanup providers, keys, local models
- Dictionary — deterministic local corrections
- History — session-only copy, re-paste, and retry
Changes autosave. Diagnostics: %APPDATA%\Undertone\app.log.
Requires Windows, Node.js 22.12+, npm, and the Windows C# compiler used by
electron\scripts\build-native.ps1.
run.batrun.bat builds and launches from source. build.bat produces the NSIS
installer. For package and smoke commands, see AGENTS.md.
Undertone can use NVIDIA Nemotron Streaming 0.6B as one local model for both live preview and the completed transcript. Choose Nemotron under Settings → Speech & AI → Local transcription engine, then install the recommended runtime. Undertone recommends NVIDIA when it detects a supported GPU and sufficiently recent normal NVIDIA graphics driver; otherwise it recommends CPU. Either choice can be overridden.
Undertone downloads versioned, size- and SHA-256-pinned runtime and model artifacts. Users do not need the CUDA Toolkit, Visual Studio, CMake, or other developer tools. Runtime packages and their provenance manifest are published in the Undertone Nemotron runtime release. The model is provided under the NVIDIA Open Model License. Whisper and Nemotron are never loaded as transcription engines at the same time. Local live preview is Nemotron-only; Whisper Large V3 Turbo remains available for completed, non-live recordings.
Audio goes only to the selected transcription provider (local keeps it on this PC). History is in memory and discarded on exit. Pasting into elevated apps requires Undertone itself to be elevated.