Push-to-talk dictation for macOS, 100% on-device. Hold a key, speak, release, your text is transcribed locally and pasted right at your cursor. Pick your engine: Qwen3-ASR, NVIDIA Nemotron 3.5 or Voxtral Mini: all running via MLX.
🌐 talkink.app: website & one-click download
- 🔒 Local & private: your voice and text never leave your Mac. No cloud, no subscription.
- ⚡ Fast: your words land in about a second; the lightest models run 30–40× faster than real time on a MacBook Air M4.
- 🌍 Multilingual: 27 languages, picked once or auto-detected. Punctuation & capitalization included.
- ⌨️ Paste anywhere: auto-pastes at the cursor; always on the clipboard as a fallback.
- 📖 Your vocabulary: teach it your names and jargon (“Talkink”, “PostgreSQL”), fixed on-device right after transcription.
- 🙌 Hands-free: double-tap the key to keep recording, tap once to stop. Scriptable via
talkink://URLs. - 📜 History & stats: every transcription kept locally, searchable, with your words-per-minute.
- 🛟 Nothing fails silently: a visible error journal and one-click problem reports (never your transcripts).
- 🟢 Open source (MIT), 100% native Swift, no Python at runtime.
- Talkink lives in the menu bar (no Dock icon).
- Hold the push-to-talk key (Right Option ⌥ by default) → recording starts. Prefer not to hold? Double-tap the key, recording locks on until you tap again.
- Speak.
- Release → local transcription in about a second → text is pasted at your cursor (if Accessibility is granted) and copied to the clipboard.
- It's also saved to History in case you need it again.
A floating pill (NVIDIA green) shows the state: recording → transcribing → done.
The recording pill, reacting to your voice:
Requirements: Apple Silicon Mac (M1–M5), macOS 14 or later.
- Download
Talkink.dmgfrom the latest release. - Open it and drag Talkink onto the Applications folder, then eject the disk image.
- Launch Talkink from Applications. No security warning, Talkink is signed and notarized by Apple (since v0.3.0).
Or, with Homebrew:
brew install --cask hasso5703/tap/talkinkHomebrew 6+ asks you to trust third-party taps first:
brew trust hasso5703/tap, then install.
Is it safe? Yes, twice over: the download is notarized by Apple (scanned and ticketed), and Talkink is fully open source, you can read every line in this repo and build it yourself.
Coming from Söyle (Talkink's former name, ≤ v0.3.3)? Your settings and history carry over automatically. If push-to-talk stops responding, remove the old Söyle entry from System Settings → Privacy & Security → Accessibility (− button), then switch Talkink on, one time.
Two, and the onboarding flips them on with you, no “+” and no dragging:
| Permission | Why |
|---|---|
| Microphone | To hear you |
| Accessibility | Detect your push-to-talk key in any app, and paste the text at your cursor |
Talkink uses Accessibility (not Input Monitoring) for the key, so it appears in the list automatically: you just switch it on, and the grant is picked up live, with no quitting and reopening.
On the first transcription, Talkink downloads the model once (~2.5 GB for the recommended Qwen3-ASR 1.7B) with a live progress bar, interrupted downloads resume where they stopped. After that: hold Right Option ⌥, speak, release → your text appears at the cursor and on the clipboard. That's it. 🎤
git clone https://github.com/hasso5703/talkink.git
cd talkink
scripts/build_app.sh Release
open dist/Talkink.appRequires full Xcode 16+ (the Metal compiler is needed, see BUILDING.md). A locally built app is signed with your own (or an ad-hoc) identity and runs directly.
Menu bar → Open Talkink → Settings tab:
- Push-to-talk key: Right Option (default), Left Option, Right Control, or Fn / 🌐. Double-tap for hands-free (on by default) locks recording until the next tap.
- Language: Auto (detect) or one of 27 languages.
- Vocabulary: your names and jargon, written your way, with the forms they get misheard as. Applied on-device after every transcription; real words are never touched.
- Model: a catalog of seven (Qwen3-ASR 1.7B & 0.6B, NVIDIA Nemotron 3.5, Voxtral Mini 4B; 8-bit / bf16 / 4-bit variants) with real size, quality and speed ratings. Download several, switch instantly, delete what you don't use.
- Auto-paste at cursor, feedback sounds, launch at login, check for updates, URL automation (off by default).
- Support: recent issues at a glance and a one-click, fully transparent Report a Problem (environment + error journal; never your transcripts).
With Allow URL automation enabled (Settings → Behaviour), Raycast, Alfred, Apple Shortcuts or any terminal can drive dictation:
open "talkink://toggle" # start hands-free dictation / stop & paste
open "talkink://record" # start (no-op while already recording)
open "talkink://stop" # stop & paste
open "talkink://history" # open the window on History
open "talkink://settings" # open Settings
open "talkink://report" # open a problem reportDictation commands are off by default: any app could open a URL, and
starting the microphone stays your explicit choice. Window commands always
work. URL-started dictation is hands-free: tap your key (or talkink://stop)
to finish.
Talkink's transcription is 100% on-device. The only network calls are:
- First-run model download (~2.5 GB for the recommended Qwen3-ASR 1.7B; other catalog options range from 760 MB to 4.1 GB) from Hugging Face into
~/.cache/huggingface. Downloads are resumable, quit anytime, it continues where it stopped. - Update check (optional, toggle in Settings), Sparkle fetches
https://talkink.app/appcast.xml. No usage telemetry is sent, there is nothing to send it to.
Verify it yourself: while dictating, list the app's open connections:
lsof -i -a -p "$(pgrep -x Soyle)" # no output = no connectionsBlock both endpoints (huggingface.co, talkink.app) in /etc/hosts and
dictation keeps working, the models live on your disk. Problem reports open
as a GitHub issue you review first; nothing ever leaves your Mac on its own.
- Push-to-talk does nothing → grant Accessibility (System Settings → Privacy & Security → Accessibility) and switch Talkink on. It re-arms itself within a few seconds; relaunch it if the key still does nothing.
- After updating from an old version (Söyle ≤ v0.2.0), the key/auto-paste stopped working (toggles look on but do nothing) → macOS ties permissions to the app's code signature, and early builds were signed differently. In System Settings → Privacy & Security, remove the old Söyle entry from Accessibility (− button), then switch Talkink on. One time, the identity has been stable since v0.3.0.
- It stopped working after rebuilding from source → ad-hoc signatures change each build; run
scripts/dev_sign_setup.shonce to create a stable local signing identity so grants persist. - Using Fn / 🌐 as the key → set System Settings → Keyboard → "Press 🌐 to" = Do Nothing.
- Model download stalls → check your connection and
~/.cache/huggingface.
| Component | Role | License |
|---|---|---|
| Qwen3-ASR | Default model (1.7B / 0.6B, 30 languages) | Apache 2.0 |
| NVIDIA Nemotron 3.5 ASR | Lightest/fastest option (cache-aware FastConformer-RNNT, 600M) | NVIDIA model license (OpenMDW-1.1), see model card |
| Voxtral Mini | Alternative engine (4B, 13 languages, auto-detect) | Apache 2.0 |
| mlx-audio-swift | Native Swift/MLX STT implementations (Prince Canuma) | MIT |
| mlx-community | MLX-converted weights (8-bit / bf16 / 4-bit) | per model license |
| MLX | Compute on Apple Silicon (Apple) | MIT |
Each downloaded model is governed by its own license (see the model cards). By using Talkink you agree to the license of the model(s) you download.
- Live text while you speak (streaming
generateStream) - Notarized + Sparkle in-app auto-update (v0.3.0)
- Homebrew install via our tap (v0.5.0), homebrew-cask once notable enough
- Custom dictionary (proper nouns, jargon), Settings → Vocabulary
- Hands-free, double-tap the push-to-talk key, tap to stop
Talkink's code is MIT. See the table above for component and model licenses.