How to use ChatGPT voice mode for real work
Talking to the desktop app instead of typing at it: where the button is, the one setting that decides whether it feels usable, and a real proposal edited out loud.
ChatGPT voice mode lets you talk to the desktop app while it works, so you can pace the room instead of sitting at the keyboard. Open it from the voice button inside any existing chat, keep the reasoning effort low or medium so replies stay quick, and interrupt whenever you want.
What voice mode actually changes
Voice mode is a live conversation with the model instead of a typing session. You press one button, you talk, it talks back, and everything either of you says lands in the chat as text, as if you had typed it.
The difference is not input speed. It is where your body can be. You can stand up, pace, walk to the other side of the office and keep working. The keyboard stops being the thing that pins you to a chair.
There is a second effect that is harder to measure. Talking a problem through forces you to finish the thought. You hear yourself say the vague version and you correct it out loud. That is a different kind of work than composing a well-formed prompt in a box, and for some jobs it is better.
I don't have to be sitting in the chair all day typing. I can actually be up and about.
Where the button is
This runs in the ChatGPT desktop app.
There are two ways in. Inside a conversation you are already having, the voice chat button sits in the bottom right of the message box. Press it and you are in voice with all of that conversation's context intact. That is the one you will use most, because it means picking a project back up where you left it.
For something new, the bottom left has a start-a-new-voice-chat option, which opens a blank conversation in voice.
Leaving is the X button. You drop back to the normal text box, where the dictation button still works as plain transcription. Clear the field and you can go straight back into voice. It is meant to be toggled in and out of, not committed to.
Turn the reasoning effort down
This is the setting that decides whether voice mode feels good or feels broken.
Pick a lower reasoning effort — light or medium, not the top setting. A high-effort model sits and thinks before every reply, and a conversation with long gaps in it is not a conversation. You will stop using it inside a week.
The tradeoff is smaller than it sounds. A light setting still handles the overwhelming majority of what you ask in a back-and-forth: shaping an idea, tightening a plan, pulling a quote out of a document, deciding what to do next. Save the heavy reasoning for the jobs where you actually want it to grind, the deep analysis you were going to wait on anyway. For talking, fast wins.
A real run: editing a proposal out loud
Here is what the first real use looked like. A client proposal needed two things added: a new benefits section and a team calculator. Instead of sitting down to it, the work happened out loud while walking around the office.
Add the section. Now pull from the call transcript and tell me the exact words they used. Good — work those in. Now put it on screen and show me what you changed.
That last instruction is the one that makes it practical. It scrolled through the updated page and walked through its own edits, so the review was thirty seconds of looking rather than a read from the top. The pattern is the same as computer use: the agent does the work, you approve it. Voice just takes the keyboard out of the approval loop.
Interrupting is the feature
You can cut it off mid-sentence and it stops. No waiting for it to finish a paragraph you already decided you did not need.
That sounds small. It is the difference between a tool and a demo. Real thinking out loud is full of interruptions, corrections and "no, go back." A system you have to wait politely for is one you will abandon.
There is also a mute button. Press it and it stops listening while you take a phone call or answer someone in the room, then unmute and carry on in the same conversation. Without that, the first interruption of an ordinary workday would end the session.
It does the thinking work too
Voice is not only for dictating instructions. Asked how to get more out of a slide deck we had already built, it came back with three uses: run a short live workshop for business owners, cut a few slides into teaching posts that point at the workshop, and send the deck after prospect calls tied to one repetitive task the prospect had mentioned. Then it picked one and said start there.
Asked to map the workshop out, it proposed forty-five minutes with five to ten owners, each bringing one task they keep re-explaining to AI, each leaving with a first working version. Told to make it fifteen minutes instead, it tightened the whole thing without argument.
That exchange took about a minute of talking. Typed, it is four or five prompts and a lot more friction before you get to the version you actually wanted.
Where it fits and where it does not
Voice is good for shaping, directing and reviewing. It is worse anywhere the exact characters matter — code, numbers, precise strings, anything you would want to check one character at a time. Type those.
Practically: use headphones. Some people run it standing at the desk. Some walk the block talking to the desktop app through their phone. The setup that gets used is the one that fits how you already think, so try it for a week before deciding.
- Start inside a conversation you already have going, not a blank one.
- Set the reasoning effort to light or medium before you start talking.
- Ask it to show you its work on screen rather than describe it.
- Interrupt early and often. That is how the back-and-forth is supposed to go.
The short version
- Voice mode lives in the ChatGPT desktop app: bottom right of an existing chat, or start a new voice chat from the bottom left.
- Keep the reasoning effort on a light or medium setting. High effort pauses before every reply and kills the conversation.
- Pair it with computer use so it can make the change and then show you the result on screen for approval.
- You can interrupt it mid-sentence and mute it to take a call, which is what makes it survive a real workday.
- Best for shaping, directing and reviewing. Type anything where the exact characters matter.
Questions this raises
Where do I find voice mode?
In the ChatGPT desktop app. Inside a conversation you are already having, the voice chat button is at the bottom right of the message box, and it carries that conversation's context in with it. To start fresh, use the start-a-new-voice-chat option at the bottom left. The X button drops you back to the text box.
Which model setting should I use for voice?
A lower reasoning effort — light or medium rather than the highest option. High-effort models pause to think before every reply, and the back-and-forth stops feeling like a conversation. A light setting still handles the large majority of what you ask it while talking. Save the top settings for deep analysis you are willing to wait on.
Can it actually change things, or does it only talk?
It can change things. In the run above it edited a client proposal, pulled specific wording out of a call transcript, and then displayed the updated page on screen so the edits could be reviewed. That is voice layered on top of computer use: the model drives the machine and you approve the result before it goes anywhere.