VibeToText

Local push-to-talk dictation for Windows / macOS / Linux. Hold a global hotkey, speak, release — your transcript pastes at the cursor, wherever you are (Slack, VS Code, browser, terminal, anywhere that accepts text). Inference is 100 % on-device — no API key, no network round-trip, no audio leaves your machine.

Download the latest release →

---

Install

Windows

1. Grab VibeToText_<version>_x64-setup.exe (NSIS installer, smaller) or VibeToText_<version>_x64_en-US.msi (MSI for managed installs) from the latest release.
2. Run the installer. It installs to %LOCALAPPDATA%\Programs\VibeToText\ (current-user, no admin prompt).
3. Launch VibeToText from the Start menu. The settings window opens; minimize it — the app lives in the system tray.
4. Press your push-to-talk hotkey (default Ctrl+Alt+D), speak, release. The transcript pastes at your cursor.

macOS

1. Grab VibeToText_<version>_aarch64.dmg from the latest release. The arm64 build runs cleanly on Apple Silicon natively and on Intel Macs under Rosetta 2.
2. Mount the DMG and drag VibeToText.app to /Applications.
3. First launch: right-click the app → Open → confirm in the dialog. macOS Gatekeeper blocks unsigned apps on double-click; the right-click path adds it to your allowlist. (We don't yet sign + notarize — see #issue if this is a blocker for you.)
4. Grant Microphone and Accessibility permissions when macOS prompts (System Settings → Privacy & Security). The Accessibility permission is what lets VibeToText paste at your cursor.
5. Press the hotkey, speak, release.

Linux

# Debian / Ubuntu / Mint / Pop!_OS / etc.
sudo dpkg -i vibe-to-text_<version>_amd64.deb
# Resolves any missing system deps:
sudo apt -f install

Then launch VibeToText from your app launcher and press the hotkey.

The .deb declares libwebkit2gtk-4.1-0 + libgtk-3-0 as dependencies; both are present on most desktops out of the box.

Other distros: extract the .deb with ar x (it's a tar archive under the hood) or build from source per BUILDING.md.

---

First-run experience

Step	What happens
First launch	Settings window opens. Configure your hotkeys, choose Auto / CPU-only mode.
First dictation press	If you're on CPU mode, the ~250 MB Moonshine model downloads (one-time, cached afterwards). On GPU mode, the ~770 MB Whisper medium.en model downloads from HuggingFace.
Subsequent presses	Hotkey-to-paste latency is 300–500 ms on GPU, 0.7–1.5 s on CPU (Moonshine). Both well under most users' thinking pauses.

Models live under:

• Whisper: ~/.cache/huggingface/hub/
• Moonshine: <app_data_dir>/models/moonshine-base-en-int8/
  • Windows: %APPDATA%\dev.vibetotext.app\models\
  • macOS: ~/Library/Application Support/dev.vibetotext.app/models/
  • Linux: ~/.local/share/dev.vibetotext.app/models/

You can pre-download from the Settings → Model files → Download model button if you'd rather not wait on the first dictation.

---

Hotkeys

Action	Default
Show / hide settings	Ctrl+Alt+V
Push-to-talk dictation (hold)	Ctrl+Alt+D
Toggle STT on / off	(unset; configurable)

All three are configurable in Settings. Modifier-only combos like Ctrl+Shift are supported on Windows via a Raw Input hook (the standard RegisterHotKey API can't reliably deliver modifier-only events).

---

What's under the hood

Two STT backends, picked at dictation time based on your "Compute device" setting and runtime CUDA detection:

• Whisper via CTranslate2
  • ct2rs — the same engine that powers faster-whisper. Default on GPU. Custom-vocabulary bias via Whisper's initial_prompt. Hallucination filter for training-set residue ([BLANK_AUDIO], "Thanks for watching", etc.).
• Moonshine base-en (INT8) via sherpa-onnx — RTFx 25–40× on AVX2 CPUs at ~6.65 % WER. Default on CPU. Faster than Whisper small.en on the same hardware at comparable accuracy.

Energy-based VAD trims leading + trailing silence before encoding. Cross-platform CUDA detection via dynamic loading — same binary works with or without an NVIDIA card.

Output goes through a clipboard-restore paste: stash existing clipboard → write transcript → send Ctrl+V → restore previous clipboard contents. Fast on long transcripts, doesn't lose your copied text.

For the full architectural detail see BUILDING.md.

---

Privacy

Everything runs locally:

• Audio is captured by cpal, resampled to 16 kHz mono, fed to the model in-process. It's never written to disk; never sent over the network.
• Models download once from HuggingFace (Whisper) or the sherpa-onnx GitHub releases page (Moonshine). After that, no further network traffic happens during dictation.
• The only data persisted is analytics.json (per-utterance counts + short transcript previews, used by the dashboard widgets). Reset any time from Dashboard → "Reset analytics".

---

Building from source

If you want to modify VibeToText or build a custom binary, see BUILDING.md. End users don't need any of this — the installer is self-contained.

---

License

MIT — do whatever.