How to apply this guide
Separate creating a transcript from sending a request to a provider or taking a device action.
Start with
Use plain local recognition for a test note. Verify what the app shows for its model and mode. If you select Assistant, check its configured writing provider before speaking anything private.
Check before you finish
Choose only the features you want. Avoid claiming that every stage is offline just because a speech model is downloaded.
Evidence to keep: Check which processing path is selected and whether it sends data off-device before using private material. This page’s key limit is: Separate creating a transcript from sending a request to a provider or taking a device action.
Start here
The same sentence can become a note, a query or an instruction. What happens to your audio and text depends on the mode and destination, not only the microphone icon.
Start by identifying the stage
Use plain local recognition for a test note. Verify what the app shows for its model and mode. If you select Assistant, check its configured writing provider before speaking anything private.
Trace what crosses the boundary
A local recognizer can process audio on-device. A separate provider may receive the text when you request rewriting or an answer. Pasting into an online app creates another destination for the finished text.
Review settings for the actual workflow
Choose only the features you want. Avoid claiming that every stage is offline just because a speech model is downloaded.




