Skip to main content
You dictated “what time is the meeting tomorrow” and instead of those words appearing, you got an answer. Or you dictated a paragraph and what landed was a tidied-up rewrite of it rather than what you actually said. Three different features can process your speech, and each one has its own switch. Working out which one is running takes a minute.

1. The voice agent

OpenWhispr has a voice agent: say its name and the rest of what you say is treated as an instruction to carry out, not words to type. The app’s own description is “When you say ‘Hey [agent name]’ followed by an instruction, the AI executes your command — composing content, formatting text, or answering questions — instead of transcribing.” So a dictation that begins with the agent’s name is routed to the agent. If your agent is called something ordinary, you can trigger it by accident just by saying that word. Where to check: SettingsLanguage Models under AI Models, then the Voice Agent tab. Enable voice agent turns it off entirely, and Agent Name is the word that triggers it.
If you want to keep the agent but stop the accidents, rename it to something you’d never say by chance.

2. The Voice Agent hotkey

There’s a separate hotkey for the agent, described in the app as “Speak a request — your AI agent types the result, not your words.” If you used that shortcut instead of your dictation shortcut, you’ll get an answer every time. Turning off Enable voice agent disables this hotkey too — they’re the same feature — so if you’ve done that and still get answers, the cause is something else on this page, not the toggle. Where to check: SettingsHotkeys under App. Compare Dictation Hotkey with Voice Agent Hotkey and make sure they’re not similar enough to hit the wrong one.

3. Dictation cleanup

Separately from the agent, OpenWhispr can run your transcription through a language model to tidy it — punctuation, filler words, obvious mis-hearings. This is Dictation Cleanup, and it’s what makes text come out polished rather than verbatim. Cleanup shouldn’t answer questions. But it is the reason your words can come back reworded, and if you want the raw transcription this is the switch. Where to check: SettingsLanguage Models under AI Models, then the Dictation Cleanup tab.
You can always see what was transcribed before any processing. In your history, each item offers View raw transcript and Copy raw transcript. Comparing the two tells you immediately whether the change happened during transcription or afterwards.

If you turned the agent off and it still happens

That’s worth telling us about, and it’s the version of this we most want to see. Check in this order:
1

Confirm which hotkey you pressed

Turning off Enable voice agent disables the Voice Agent hotkey too, so if you’re still getting answers with the toggle off, double-check you pressed the dictation hotkey and not the agent one.
2

Compare the raw transcript

If the raw transcript is your actual words and the final text is an answer, something after transcription rewrote it. If the raw transcript is already an answer, the request went to the agent.
3

Check any custom prompt

Agent-name detection itself runs directly against your transcript, before any prompt is involved — see how naming it works. The placeholder warning in Prompt Studio is about something else: a custom agent or cleanup prompt that drops the agent-name placeholder can still behave unpredictably once the agent is triggered, so keep it if you’ve edited that prompt.
4

Send us a debug log

It records which path the request took. That turns this from guesswork into a single answer — see how to send one.

FAQ

Turn off Dictation Cleanup and Enable voice agent, both under SettingsLanguage Models under AI Models. Transcription then gives you what you said.
Yes — the voice agent can run on OpenWhispr Cloud, your own API key, a local model, a self-hosted endpoint, or an enterprise provider. The mode is chosen on the Voice Agent tab.
Because the trigger is what you said. A dictation that happens to start with the agent’s name is routed to the agent; one that doesn’t, isn’t.