Offline dictation for Mac private, on-device, good with code

Option 1: macOS Dictation

Every Mac has dictation built in.

1. Open System Settings → Keyboard and turn on Dictation.
2. Pick a Shortcut, for example pressing the Globe key or Control twice.
3. Click into a text field, press the shortcut and speak.

On Apple Silicon, Apple says dictation requests are processed on device for many languages; some languages and features may still use Apple's servers, and Apple documents which in its dictation help. It costs nothing and needs no install.

For coding, the gaps are vocabulary and control. There is no list where you teach it your repo names, library names or teammates' names, and you cannot choose the model.

Option 2: Voice Control

System Settings → Accessibility → Voice Control is Apple's full hands-free control system. It runs on device and has a Vocabulary list for custom words. It is built for controlling the whole Mac by voice, so it listens for commands all the time it is on and takes some learning. It is a good fit if you want hands-free use; it is heavy if you only want to dictate a paragraph now and then.

Option 3: apps with local speech models

Open speech models now run fast on Apple Silicon. OpenAI's Whisper and NVIDIA's Parakeet are the common ones. Apps built on them download the model once and transcribe on your Mac from then on. When you compare them, look at:

• Where it runs. Local by default, with any cloud step clearly optional.
• Push-to-talk in any app. Hold a key, speak, release, and the text lands at the cursor, including in terminals and editors.
• Custom vocabulary. A way to add names and jargon so pnpm does not come out as "PNPM" or "p and p.m."
• Latency. How long after you let go the text appears. Anything over a second breaks your flow.
• Silence handling. Some models invent words when they hear only a fan or keyboard noise.
• Cleanup. Punctuation and filler-word removal, ideally local too, with a way to turn it off for terminals.

How to check it is offline

Turn off Wi-Fi, unplug Ethernet, and dictate a sentence. If text still appears, transcription is local. Then read the app's privacy page for what it sends when you are online; downloading a model or checking for updates needs the network even when transcription does not.

Dictation in Cadence

Cadence is a free macOS app for running coding agents, with on-device dictation built in. Dictation works without any agent setup.

Set it up

1. Download Cadence and grant Microphone, Accessibility and Input Monitoring during onboarding.
2. Choose your hotkey. Option (⌥) is recommended; change it later in Settings → Hotkey → Push-to-Talk Key.
3. Download the speech model when asked: Parakeet v2, about 630 MB. After that, transcription needs no network.

Use it

Click into any text field, hold ⌥, speak, and let go. The text is pasted at your cursor and your previous clipboard is put back. A quick tap, or a shortcut like ⌥←, types nothing. For long passages, click the mic button on the toolbar to dictate continuously.

Cadence keeps the half second before you press the key so your first word is not clipped, finds speech with Silero VAD, and transcribes each piece at the next pause while you are still holding the key. On Cadence's own test set this measured 1.88% word error rate on clean read speech and about 330 ms from release to text at the 95th percentile. For other languages, pick Parakeet v3 (25 languages) in Settings → Models.

Teach it your words

Add names and jargon in Settings → Commands → Vocabulary; the model is primed with them, and your agents' names are added automatically. With a Whisper Large v3 Turbo model downloaded, Cadence double-checks words that might be from your vocabulary. If a phrase is always heard wrong, add a rule in Settings → Commands → Auto-Corrections.

Text Polish

Text Polish fixes punctuation, capitalization and filler words with Qwen 2.5 1.5B running on your Mac (a separate download of about 1.1 GB, in Settings → Injection → Text Polish). Set a tone per app in Settings → AI → App Tone Profiles; use Raw (no polish) for terminals. Cloud Polish is optional, off by default, and sends text (never audio) to Anthropic with your own API key.

What stays on your Mac

Audio, transcription and Text Polish run locally, and audio is not uploaded. The microphone stays open while Cadence runs so the first word is not cut off; that half second lives in memory and is thrown away unless you press the hotkey. Details are on the privacy page.

The same key also talks to your coding agents: a double-tap of ⌥ opens Talk instead of dictating. Plain dictation never goes to an agent.

Common questions