“Local” is a useful word, but it is not a complete explanation. A voice workflow usually has several steps: capture audio, recognize speech, clean up the transcript, translate or rewrite it, save a result, and sometimes send something to another service.
When I want to know whether a voice app is private, I do not start with the biggest promise on the homepage. I ask which model handled each step.
Start with the model path
In Voxt, an ASR channel handles speech recognition and an LLM channel can handle cleanup, translation, or rewriting. If both are local and the required models are installed, those model steps happen on the Mac.
That is a much more useful statement than “the app is private.” It tells you what local processing covers, and it leaves room for the parts of the product that are not model inference: downloads, accounts, billing, analytics, support, and updates.
Local does not mean carefree
Keeping audio on the Mac removes one kind of exposure: a hosted inference service does not need to receive the content for the local model step. It does not remove the need to protect the Mac itself.
The local boundary includes recordings, transcripts, model files, history, backups, user permissions, and anyone who can log into the device. A locked Mac with sensible disk and account settings is part of a private voice setup. So is checking whether a meeting recording is still sitting in a folder after you are finished with it.
Remote can be a deliberate choice
Sometimes a remote provider is the right tool. You may need a particular language, a larger model, lower local resource use, or a quality level your current Mac cannot provide. BYOK makes that choice explicit: you connect the provider and key, and the data needed for that step follows the provider's policy.
The important thing is that remote processing is not confused with local processing. A sensible setup might use local ASR for a sensitive voice note and a remote model for a non-sensitive rewrite. Or it might keep a fully local path for travel and switch providers later.
A small checklist before a sensitive recording
Before using voice input for confidential material, I would check:
- Which ASR channel is selected?
- Which LLM or translation channel is selected?
- Are the needed local models already installed?
- Where will the recording and transcript be stored?
- Is any remote provider enabled for this task?
- Have the Mac's microphone and Accessibility permissions been reviewed?
None of this is complicated. It is just more honest than treating a single privacy label as a substitute for knowing the path.
For Voxt, the Privacy & Data Flow page is the best place to check the current boundaries. Local-first is a configuration choice that gives you control; it is not a claim that every feature, service, or provider is offline.