In this guide
The quick answer
Echo’s speech recognition runs on your Mac. After the speech model is downloaded, choose Off for the simplest offline setup. Apple cleanup and a downloaded model running locally can also work offline. Cloud AI cleanup needs a connection.
Offline dictation has two stages
First, a speech model turns audio into text. Then, if enabled, a cleanup model edits that text. A tool can run the first stage locally while sending the second to an online service. Check both before relying on it without a connection.
Echo’s main speech model runs on your Mac after a one-time download of about 630 MB. The choice in Settings → Polish determines what happens next. You do not need an Echo account.
| Mode | Before going offline | What it does |
|---|---|---|
| Off | Download the speech model. | Removes selected fillers and applies your dictionary. No cleanup model needed. |
| Apple | Make sure Apple Intelligence is available and its on-device model is ready. | Uses Apple’s on-device model to tidy text. |
| AI → Local model | Download a model and start a local server such as Ollama or LM Studio. | Uses a model running on this Mac. Keep the server running. |
| Cloud AI | An internet connection and the provider’s account are needed. | Not an offline cleanup option. Choose another mode before disconnecting. |
Prepare while you still have Wi-Fi
- Install Echo on a supported Mac: macOS 26 with Apple silicon.
- Complete setup and let the speech model download finish. In Settings → Speech, check the model’s state.
- Select the language you plan to use. If it uses an Apple speech model, let that model finish preparing too.
- Choose Off, Apple, or a local AI model in Settings → Polish. For a first offline test, Off has the fewest dependencies.
- Open a local TextEdit document and dictate a short paragraph. Resolve any microphone or Accessibility permission requests now.
A separate local cleanup model can take substantially more disk space than Echo’s speech model. Finish that download and test it before travelling; the small Echo app download alone does not include every model.
Do a real disconnected test
Turn off Wi-Fi and disconnect any other network connection. Keep your local cleanup server running if you chose one. In TextEdit, speak a fresh sentence, release the key, and check that the text appears.
Try the language and mode you actually intend to use. A successful English test in Off mode does not check a different language’s model or a local AI server. For multilingual work, repeat with each language you need.
If recognition works but cleanup waits or fails, check whether you left a cloud provider selected. If no text appears, use the listening and pasting checks. Reconnect afterwards if you still need to download something.
Local speech does not make every destination local
Echo does not store dictation audio. Its optional history contains text on your Mac, and you can clear it or disable it. Read Echo’s privacy details for the difference between speech, cleanup, and history.
The app receiving your text has its own behaviour. A cloud document, synced note, or email draft may upload that text when connected. For a completely local draft, use a local file outside a synced folder and a local cleanup mode.
Offline dictation is most useful when the whole workflow is ready: the right models, a working microphone, the chosen cleanup mode, and a place to save your words.