Whether a dictation app holds your microphone open when you are not dictating is not a privacy policy question — it is a design question, decided by how the app works out when a sentence starts and ends. There are three models in common use, and they behave very differently while you are not speaking.
The three listening models
Push-to-talk. The microphone is open while a key is physically held down and closed the rest of the time. The listening window is a physical action, so there is no state to forget about and nothing to leave running by accident. The full definition covers where the term comes from.
Toggle. A shortcut switches listening on, and it stays on until the same shortcut switches it off. The microphone is open for as long as you forget, which on a bad day is the whole meeting that started right after.
Voice activation. The app listens continuously and uses a pause in speech to guess that you have finished. The microphone is by definition open when you are not dictating, because that is the only way it can notice that you have started.
A wake word is a fourth case and behaves like voice activation: something must be listening in order to hear the word.
What macOS tells you regardless of the app
You do not have to take any app's word for it. macOS shows an orange dot in the menu bar whenever the microphone is in use, and clicking Control Centre names the application holding it. That indicator is the system's, drawn above every app, so it is the honest answer to "is anything listening right now".
Watch it for a minute while not dictating. If it is lit, something is holding the microphone open.
Open microphone and uploaded audio are separate questions
These get conflated and they should not be. An app can hold the microphone open all day and process everything locally. An app can open it for two seconds and send those two seconds to a server. The questions to ask are:
- When is the microphone open? The three models above.
- Where does the audio go? On-device processing, or uploaded to a service.
- Is anything kept? Audio, transcripts, or neither.
Private speech-to-text on a Mac works through the second and third of those across the tools people actually use.
How Coii VoiceInput behaves
Sending the revised figures over now, with the caveat that the third column is still provisional.
Hold is the default mode: the microphone is open while ⌥ Space is held down and closed when you let go. There is no silence timeout to tune, no wake word, and nothing listening between dictations. A toggle mode exists for people who want it — holding a key for a two-minute paragraph is a real problem if your hands are the reason you are dictating — and choosing it is a decision you make in settings, not a default you inherit.
Hold is the default because it makes the end of an utterance unambiguous. You stop speaking and let go; there is no silence threshold that can be wrong about whether you were finished or just thinking.
The audio never leaves the Mac, and neither do the transcripts. Push-to-talk dictation on a Mac covers what that model costs as well as what it gives you, which mostly comes down to whether you have a hand free.