How is this different from ChatGPT's voice mode?

ChatGPT's spoken conversation features and Keebye text dictation solve different jobs. Keebye puts editable text in the prompt box so you can inspect it before sending; check OpenAI's documentation for the current behavior of ChatGPT's own voice features.

Does OpenAI ever hear my audio?

No. Speech-to-text runs on-device and audio never leaves your Mac. ChatGPT receives only the final text you choose to send, exactly as if you had typed it.

Does it work in the browser and the desktop app?

Keebye inserts into ordinary editable fields in browsers and desktop apps without a ChatGPT-specific plugin. Secure fields are deliberately refused.

Does it work offline?

Dictation does, after the one-time model download — transcription needs no network connection. ChatGPT itself of course needs one.

How do I start and stop dictating?

Hold Right Command as you dictate the prompt and release when done; a tap toggles for longer sessions, and Esc bails out. If Right Command is inconvenient, Fn and Right Option are the built-in options.

Voice Dictation for ChatGPT on Mac

Speak long prompts into ChatGPT in a browser or the desktop app. On-device transcription keeps the audio on your Mac.

Why dictate prompts to ChatGPT

Suppose you need help reviewing a launch plan containing customer context, three constraints, and one decision you have not made. You want to say the full thought, inspect the wording, remove anything too sensitive, and only then choose Send. That is a text-dictation job, not a hands-free conversation.

ChatGPT's own spoken features handle a different interaction. Keebye is input dictation: your words become text in the prompt box, where you can read them, remove private detail, and add to them before anything is sent. The same hotkey also covers your editor and your terminal without a per-tool plugin.

How it works

Click into the ChatGPT prompt box in your browser or the desktop app, hold Right ⌘, and say the prompt. On release, your words appear at the cursor for you to review, edit, or send. Keebye inserts into the focused editable field, so there's no ChatGPT-specific setup.

The transcription engine lives on your Mac: Parakeet covers English out of the box, and once the model is downloaded, no connection is required and no telemetry is sent. Your audio never goes to OpenAI or anyone else; ChatGPT receives only the final text you decide to send.

Setup in two minutes

Two minutes covers it: install, say yes to the Accessibility and microphone permission prompts, and settle on a hotkey — Right ⌘ works out of the box, with Fn and Right ⌥ as the other choices. Add launch-at-login and it's waiting whenever a browser tab is.

Then add recurring product names, project names, and technical terms to the custom dictionary to reduce repeated corrections. Review proper nouns before sending.

Limits, honestly

The text arrives after you release the key, in one batch — not as live captions while you talk, and not as a spoken back-and-forth with the model. For an actual voice conversation, OpenAI's own voice mode is the tool.

Keebye is macOS only.

Now finish the opening job: dictate the launch context, inspect the text, remove what the model does not need, and send only when it reads correctly. The reason that review step matters is explored in Dictation without the AI rewrite. For the local-versus-cloud decision, read Keebye vs Wispr Flow.