How to use ChatGPT at work without typing (voice method)

Typing is not the only way to talk to AI.

I say this in almost every class. Most people still sit at a desk, type a prompt, wait, type again.

I don't like sitting still in front of a flat screen. My head feels shut in. For years I assumed that was my ADHD.

Then I looked at the research.

Why voice beats typing for thinking

Two reasons.

Speed. When I type a prompt, I manage maybe 50 to 80 characters a minute. When I talk, I get through well over a hundred in the same minute. More of what is in my head reaches the AI.

Movement. A Stanford study in 2014 found that people's creative output rose by an average of 60 percent while walking, compared with sitting (Oppezzo and Schwartz, Stanford News, 24 April 2014).

So I stopped typing my first drafts.

七步成思: think in seven steps

In Chinese there is an old story about composing a poem in seven steps, 七步成詩. I changed one character. 詩, poem, becomes 思, thought.

七步成思 just means this: walk while you think, and think out loud to the AI.

At home I keep a small box in the bathroom and another in the living room. I put my phone in the box, open ChatGPT, start a voice conversation, and talk while I get ready or pace around.

A real example. I have to send a proposal at nine. It's half past eight and I haven't showered. So I talk it through in the shower: what the client asked, what I want to say, how the deck should run. By the time I'm dressed, the outline is there.

That box goes with me when I travel. It's a habit now.

The method, step by step

1. Set your language before you start. In ChatGPT, open Settings, then Voice, then Language. Pick the language you actually speak. ChatGPT's help page says this helps it understand your speech more accurately. I speak Cantonese mixed with English; I set it, then just talk.

2. Tell it to wait. The biggest problem with voice is that the AI jumps in every time you pause. Start with: "Wait until I ask you to respond." OpenAI's own help page suggests exactly this. Long pauses or background noise can still trigger it, so a quiet room helps.

3. Talk for ten minutes. Don't edit yourself. Say what the job is, who it's for, what you already know, what worries you. Messy is fine. This is the collecting stage.

4. Then ask for one thing. "Now give me a five-point outline for the proposal." Or: "Ask me the three questions you'd need answered before writing this."

5. Finish in text. When you end the voice conversation, the transcript is added to the chat. Open it on your laptop and ask for the actual deliverable there: the email, the outline, the table.

One warning from OpenAI's help page: voice transcripts are not verbatim records. If you need an exact record of what you said, use Dictation instead, which lets you review and edit the text before sending.

My extra step: I also record my side of the conversation with the phone's recorder, then give the recording to Gemini or NotebookLM for a clean transcript. The ideas I speak are often better than what the chat summary keeps.

What's new in ChatGPT Voice (October 2026)

As of 16 October 2026, OpenAI's help page lists three voice options under Settings, then Voice:

Live: the newest, powered by GPT-Live-1.1. It can listen and speak at the same time, so you can interrupt it. It can search the web and use memory.
Advanced: the previous real-time voice. Use it on mobile if you need video or screen sharing, which Live does not support.
Standard: turn by turn. It transcribes your speech first, then answers.

Two settings worth turning on for walk-and-talk:

Background conversations: keep talking with your phone locked or in another app.
Start with Voice: opening a new chat starts voice automatically.

Voice time is limited by plan, measured over a rolling 24 hours. The page lists, for example, limited access on Free, 3 hours on Go and Plus, and 15 hours on the US$100 Pro plan. Check the current table before you rely on it.

What to keep out of a voice chat

Talking is so easy that people say things they would never type.

Don't say client names, staff names or numbers you wouldn't paste into a public tool.
OpenAI's page says audio clips from Live and Advanced conversations are stored with the transcript and kept for 30 days. Deleting the chat also deletes the clips within 30 days, with some exceptions.
Check Settings, then Data Controls, to see whether "Improve the model for everyone" is on for your account.

If you need a quick rule for what is safe to say, use my traffic-light test.

A note for Hong Kong

Hong Kong is not on OpenAI's list of supported countries and regions for ChatGPT (checked 16 October 2026). Before your team builds a habit on it, read which AI a Hong Kong team can actually use at work. The method itself is not tied to one tool. Any AI with a good voice mode works the same way: talk first, finish in text.

FAQ

Is voice slower because I have to fix the transcript?

No, because you don't fix the transcript. You use it as raw material, then ask for the finished piece in text.

Can I speak Cantonese to ChatGPT?

You can choose your language in Settings, then Voice, and ask it to speak a different language mid-conversation. Test it with your own accent before relying on it.

When should I use Dictation instead?

When you need the exact words: a quote, a message you will send as written. Voice is for thinking out loud; Dictation is for a precise prompt.

Sam Wong (黃力桐) is a Hong Kong AI adoption trainer who has trained 11,000+ people.