Dictation that sends the message.
Hold fn, talk, let go. Your words land at the cursor, exactly as you said them. In Claude, Codex or ChatGPT, press return mid-take and Saylark sends it. Free, open source, and the speech recognition runs on your Mac.
How it works
Hold, talk, release.
No window to open and no button to click. It works in any app with a text field.
Hold fn
Or Right Option, or Right Command. A small pill appears at the bottom of the screen.
Talk
Say it the way you would say it to a person. The pill moves with your voice.
Let go
Your words are pasted at the cursor, and your clipboard is put back the way it was.
Hands-freeDouble-tap the key for a long take, then tap once to stop.
Return sendsPress return mid-take to stop, paste and send in one go.
Changed your mind?esc cancels. The take is kept in case you want it later.
Return sends
Press Return and it sends.
Talking to a coding agent? Say the whole prompt, then press return before you let go. Saylark stops recording, pastes your words into the chat box and presses Return for you. One take, one message.
- On by default in three apps. The Claude, Codex and ChatGPT desktop apps. Add any other app in Settings, or turn it on everywhere.
- Hands-free for long prompts. Double-tap the key and talk for as long as you need, then press return to send.
- Changed your mind? esc cancels instead. Nothing is pasted, and the take is kept in case you want it later.
Literal text
Your exact words.
Saylark has no AI step that rewrites, summarises or polishes what you said. When you are giving instructions to an agent, a rewrite you didn’t ask for can change what you meant, so the text is your words as spoken.
You say
Um, so the the login test fails on Safari new line check the auth middleware firstSaylark types
So the login test fails on Safari
Check the auth middleware first.- Removed: fillers like um, uh, er and hmm, and stutter repeats like “the the”.
- Spoken commands: “new line” and “new paragraph” become line breaks.
- Left alone: everything else, including numbers, units, file names and identifiers. Your dictionary fixes the spelling of names.
Never lose a take
Audio is saved while you talk.
Saylark writes your audio to disk as you speak, so a long take survives a crash. If a take is cancelled or fails, it is still there to recover.
- Cancelled and failed takes stay. Paste them or transcribe them again from Home.
- No text field? A card holds your words. They are already copied, so press ⌘V anywhere, or choose “Paste when I click a field”.
- Every take in one place. Copy, paste again or play back what you said. Audio is kept 30 days by default; you can change that.
Your dictionary
A dictionary that’s just a text file.
Teach Saylark the names, products and jargon you say every day. They live in one plain text file: edit it on the Dictionary tab, in any editor, or from the command line. Changes apply within a second.
dictionary.txt
# A term on its own: spell it this way.
kubectl
# Or list what the model writes instead.
PostgreSQL: post gress
Next.js: next justOr let your own agent keep it current
saylark dict add Supabase --as "super base"
saylark dict listFor example, a daily agent job that adds the new names from your commit messages. Saylark itself never reads your messages.
Privacy
What leaves your Mac: model downloads, nothing else.
Saylark recognises speech on your Mac’s Neural Engine with Parakeet, an open speech model. The app’s only network use is downloading that model once from Hugging Face (about 600 MB), plus a small 100 MB spelling model if you use the dictionary. There is no Saylark server, no account and no telemetry in the app.
Your words aren’t sent anywhere
Audio, text and history live in your user folder. Once the model is downloaded, dictation works with Wi-Fi off.
No account, no subscription
Download it and use it. No sign-up, no word limits, no paid tier.
Open source
MIT licensed. Read the code, build it yourself, change what you like.
The Saylark window
Your takes, your words, your settings.
Home lists what you said, with words today and this week. Open it from the menu bar whenever you want it, and it stays out of the Dock the rest of the time.
Install
Drag to Applications, grant four permissions, talk.
Requirements
- Mac
- Apple Silicon (M1 or later)
- System
- macOS 26 or later
- First launch
- Downloads the speech model once from Hugging Face, about 600 MB
- Price
- Free, MIT licensed
- Download Saylark.dmg and open it.
- Drag Saylark to Applications, then open it from there.
- Grant four permissions. A short setup walks you through each one and says why.MicrophoneInput MonitoringAccessibilityPaste from Other Apps
- Try it. Hold fn and say something in the practice field. When macOS asks about pasting from other apps, choose “Always Allow”: Saylark reads the clipboard only to put your own contents back after a paste.
FAQ
Questions
How is this different from other local dictation apps?
Four things. Return sends: in the Claude, Codex and ChatGPT apps (or any app you add), press Return mid-take and Saylark stops, pastes and sends. Literal text: there is no AI rewriting step; Saylark removes fillers like um and uh and stutter repeats, and types everything else as you said it. Takes are never lost: audio is written to disk while you talk, and cancelled or failed takes can be recovered from Home. The dictionary is a plain text file with a small command line tool, so your own agent can keep it up to date.
Is it really free?
Yes. Saylark is open source under the MIT licence. There is no account, no subscription, no trial and no paid tier.
Does anything I say leave my Mac?
No. Speech recognition runs on your Mac, on the Neural Engine. Your audio, your text and your dictionary stay in your user folder. The app uses the network only to download its speech models from Hugging Face: about 600 MB on first launch, plus a 100 MB spelling model once your dictionary has terms. The app has no telemetry. This website counts page visits with Umami, without cookies or personal data.
Does Saylark rewrite or tidy up what I say?
No. It removes fillers (um, uh, er, hmm) and stutter repeats like “the the”, turns “new line” and “new paragraph” into line breaks, and applies your dictionary. Everything else is the speech model’s transcript of your words, as you said them.
Which Macs does it run on?
Macs with Apple Silicon (M1 or later) running macOS 26 or later. The first launch downloads the speech model, about 600 MB.
Can I use a key other than Fn?
Yes. In Settings you can choose Fn, Right Option or Right Command. Hold it to talk, or double-tap it for hands-free and tap again to stop. Esc cancels.
What if I start talking with no text field selected?
Your words are copied and a small card appears. Press ⌘V anywhere, or choose "Paste when I click a field" and click where the text should go.
Why does it need Accessibility and Input Monitoring?
Input Monitoring lets Saylark notice when you hold the trigger key. Accessibility lets it paste at your cursor and press Return for you. Microphone is for your voice. macOS also asks whether Saylark may paste from other apps: choose Always Allow, because Saylark reads the clipboard only to put your own contents back after it pastes.
I use another dictation app. Will they clash?
If both listen to Fn, both will paste. Quit the other app, or pick a different trigger key in Saylark’s Settings.
Which languages does it understand?
Saylark is tuned for English today.
ShipTiffin