Guide
Dictating into ChatGPT
Good prompts are long, and long prompts are faster spoken. The question is whose machine does the listening.
The short answer
Voz transcribes on your Mac and types the result into the prompt box, so ChatGPT receives text you already have rather than audio it has to hear.
It works in the ChatGPT desktop app, in a browser tab, and in every other app on the same hotkey — because it types wherever the cursor is rather than integrating with anything.
ChatGPT already has a microphone button. Why use this instead?
Because they do different things. ChatGPT's dictation and Voice Mode send your audio to OpenAI to be transcribed there. That is fine for most prompts and it is how the feature works — but it means the recording of your voice leaves your machine, and it only works inside ChatGPT.
Voz transcribes locally and hands ChatGPT ordinary text. Two consequences worth having: the audio never exists anywhere but your Mac, and the same key works in Slack, your editor, and everything else. One habit instead of one button per app.
If you want a conversation rather than a prompt, use Voice Mode. It is genuinely better at that, and we are not pretending otherwise.
Setting it up
- Install Voz and grant microphone and accessibility permission when macOS asks. Accessibility is what lets text land at the cursor.
- Click into the ChatGPT prompt box, in the app or the browser.
- Hold your push-to-talk key, say the whole prompt, release.
- Read it before you send. Dictation gets proper nouns wrong more often than it gets sentences wrong.
Why long prompts are the use case
Most people speak at about 150 words a minute and type at closer to 60. That gap does not matter for "summarise this" — it matters enormously for the paragraph of context that actually changes the answer.
The prompts that work are specific: what you have, what you want, what you already tried, what the constraints are. Typing that is a chore, so people skip it and get a worse answer. Speaking it costs twenty seconds.
Worth knowing
- Add the names you use constantly — libraries, colleagues, products — to custom vocabulary, so they come out spelled right.
- Say "new line" for a line break, and punctuation by name if you want it exactly.
- Paste code rather than dictating it. Prose is what speech is good at.
- If you work under an NDA or with client data, local transcription is the difference between one upload and two.