How-toMac6 min read
Offline Speech to Text for Mac: Step-by-Step Setup
Set up offline speech to text for Mac: download a local speech model, pick on-device cleanup, check each app route, and dictate with Wi-Fi off.
Published by Optimal Apps, the maker of Petal.
On this page
To get offline speech to text for Mac, you need a speech model that runs on the Mac, not on a server. Apple Dictation does this for many languages on Apple silicon. For more model choices and text cleanup that also stays local, use Petal: download a speech model one time, select Petal W1 for cleanup, check that no app route uses a cloud provider, and then dictate with Wi-Fi off. This guide shows each step.
The steps match Petal 2.8.1 on macOS 26, with the Parakeet 110M speech model and the Petal W1 cleanup model. Petal is free and open source, and Optimal Apps publishes it.
Quick checklist
| Check | Where | Why |
|---|---|---|
| A speech model is downloaded | Settings > Models | Speech recognition runs from this file on the Mac |
| Cleanup is Petal W1, Apple Intelligence, or Off | Settings > Intelligence > Text Cleanup | A remote Cloud Model needs the internet |
| No route uses a remote Cloud Model | Settings > Router | A route can override the default cleanup |
| Microphone and Accessibility are on | System Settings > Privacy & Security | Petal needs them to record and to paste |
| The Mac has Apple silicon | Apple menu > About This Mac | Petal ships for Apple silicon only |
Option 1: Apple Dictation offline
Apple Dictation is the fastest way to start, because it is built in.
- Open System Settings > Keyboard.
- Turn on Dictation.
- Pick your language and the shortcut.
- Turn off Wi-Fi and press the shortcut in a text field.
On a Mac with Apple silicon, many languages process on device. Apple's help says that dictation stops after 30 seconds without speech, and it does not work while Voice Control is on. It does not remove filler words or rewrite your text. If it fails, read Mac dictation not working.
Offline speech to text for Mac with Petal, step by step
Step 1: Install Petal
- Download the latest DMG from GitHub Releases, or use the button on the Petal for Mac page.
- Open the DMG and drag Petal to Applications.
- Open Petal. A petal icon appears in the menu bar, and the setup window opens.
Step 2: Choose and download a speech model
The setup asks you to Choose a Model. Each card shows a speed meter, a quality meter, and the download size.
| Model | Size | Languages | Use it for |
|---|---|---|---|
| Parakeet 110M (recommended) | ~455 MB | English | Fast daily dictation |
| Parakeet 0.6B V3 | ~485 MB | 25 European languages | Fast dictation in German, French, Spanish, and more |
| Parakeet Unified 0.6B | ~625 MB | English | A live transcript while you speak |
| Whisper Large V3 Turbo | ~1.1 GB | 99 languages | Accuracy over speed |
| Qwen3 ASR 1.7B | ~2.5 GB | 30 languages | The highest accuracy, slower |
Pick Parakeet 110M if you dictate in English. The download needs the internet one time. After that, the model is a file on your Mac.
You can change the model later in Settings > Models. To open Settings, click the menu bar icon and choose Petal Settings…, or press Command-Comma while the menu is open.
Step 3: Set the shortcut and permissions
- On Record a Shortcut, keep the default Option-Space or press your own.
- On Microphone Access, allow the microphone.
- On Accessibility Access, turn on Petal in System Settings > Privacy & Security > Accessibility. Petal uses this permission to paste into the front app.
Permissions are local. They do not need the internet. If the paste fails later, read accessibility permission not working on Mac.
Step 4: Select Petal W1 for cleanup
Raw transcripts contain "um", repeats, and corrections. Petal W1 is an on-device model that condenses rambles and repeats and keeps your voice. It is about 1.0 GB.
- Open Settings > Intelligence.
- Under Text Cleanup, select Petal W1.
- Wait for the download to finish. Until it finishes, Petal pastes your words as you said them.
Other offline choices:
- Apple Intelligence uses the on-device Apple model and follows your own instructions.
- Off pastes the transcript with no cleanup.
Do not pick Cloud Model for offline use, unless the server is on your Mac. A local server such as Ollama or LM Studio at http://localhost:11434/v1 also works offline. For that setup, read per-app dictation cleanup with your own AI key.
Step 5: Check the Router route for your target app
A route can replace the default cleanup for one app or website. If one route uses Cloud Model with OpenAI, that app needs the internet even when the rest of Petal is offline.
- Open Settings > Router.
- Under Routes, click each app or website.
- Under Instructions, look at the Intelligence menu. Default (Petal W1) is offline. Petal W1 and Apple Intelligence are offline. Cloud Model with a remote provider is not.
- Click All Other Apps and check it too.
Step 6: Turn off Wi-Fi
- Click Control Center in the menu bar.
- Click Wi-Fi to turn it off.
- If you use Ethernet, unplug the cable to make the test real.
Step 7: Dictate into an editor
- Open TextEdit or your code editor and click in the document.
- Press Option-Space and speak: "Um, so the meeting is on Friday, no, Thursday at ten."
- Press Option-Space again to stop. You can also hold the shortcut and release it to stop.
- Petal shows a small capsule while it works, and then pastes the text at the cursor.
With Petal W1 on, the result reads close to "The meeting is on Thursday at ten." To compare, open Settings > History. Each card shows the Transcript and the Cleaned Up text.
Why offline dictation fails, and the fix
- The download did not finish. The model card says "Not downloaded" or "Paused". Connect, then resume the download in Settings > Models.
- A route uses Cloud Model. Change its Intelligence menu to Default or Petal W1.
- The paste does not arrive. Accessibility is off or stale after an update. Turn Petal off and on in Accessibility.
- Petal W1 is not ready. The Router shows a warning: "Petal W1 is not ready." Click Set Up and finish the download.
- The Mac is an Intel Mac. Petal ships for Apple silicon only. Use Apple Dictation, or Voice Control, which Apple suggests for offline use on older Macs.
Local speech recognition on Mac: keep models up to date
Petal checks for app updates through Sparkle. An update can add new models, but your downloaded models stay on the Mac. To free disk space, right-click a model card in Settings > Models and choose Delete Download….
How to check that nothing leaves the Mac
You do not have to trust a claim. With Wi-Fi off, a dictation that pastes text proves that the path works offline. If you want more proof, use a network monitor such as Little Snitch or LuLu, and watch for connections while you dictate. Petal is open source, so you can also read the code on GitHub.
Other apps that work offline
Superwhisper, MacWhisper, VoiceInk, and Spokenly also offer local models. Wispr Flow does not: its privacy page says that transcription always happens in the cloud. If you came here from that problem, read Wispr Flow not working offline. For a full list, read the best dictation apps for Mac.
Get Petal for Mac to start.
Questions
Can I dictate on a Mac without internet?
Yes. Apple Dictation works offline for many languages on Apple silicon. Apps with local models, such as Petal, also work offline after you download the model one time.
Does Petal need the internet after setup?
No. After you download a speech model, and Petal W1 if you want cleanup, the default path runs on your Mac. Only a Cloud Model route with a remote provider needs the internet.
Which offline model should I pick in Petal?
Start with Parakeet 110M for fast English dictation. For other European languages, use Parakeet 0.6B V3. For 99 languages, use Whisper Large V3 Turbo, which is slower but more accurate.
Does Apple Intelligence cleanup work offline?
Apple Intelligence text cleanup in Petal uses the on-device Apple model, so it does not need a server. Petal W1 is the other on-device choice.
Why does my dictation still fail offline?
The usual cause is a route that uses Cloud Model with a remote provider, or a model that did not finish its download. Check Settings > Router and Settings > Models.