Story
When my hands are busy and the AI answers out loud, I want to see the words as they are spoken and cut in with my voice, so I can catch a number I missed and steer it without touching the screen.
Fatima, baking bread in her kitchen in Leicester, with her phone propped against the flour jar.
- 01 Setting
Fatima has dough on her hands. She asks out loud for the next step.
- 02 Trigger
The answer has 2 times to remember: 10 minutes and 20 minutes.
- 03 Without it
An orb (a glowing ball on the screen) talks. Halfway through kneading she forgets the time and asks again.
- 04 With it
Each word appears in the chat as it is spoken. A glance shows "10 minutes".
- 05 Outcome
She ends the voice chat with a knuckle. The steps stay in the chat for later.
Live demo
This demo follows a script, so it plays the same every time - there is no real AI behind it. Change a setting to see the same moment play out differently.
What happens, step by step
| State | The person sees | The AI is | When | Screen reader |
|---|---|---|---|---|
| 1 | The text box becomes a voice bar: "Listening", mute and End. | Listens. | 0 s | “Voice chat on. Listening” |
| 2 | Her words appear in her bubble as she says them. | Listens. | while she talks | - |
| 3 | Each word appears as it is spoken. The current word is highlighted. The bar says "Speaking". | Speaks. | about 2-3 words a second | “Speaking” |
| 4 | She speaks. The AI stops at once. Its line ends with "stopped, you spoke". | Stops and listens. | within a moment | “Listening” |
| 5 | A short answer to the new question, shown and spoken. | Speaks. | - | “Speaking” |
| 6 | The full answer as text. The bar goes back to "Listening". | Listens. | after the last word | “Listening” |
| 7 | "Voice chat ended". The words stay in the thread. The text box is back. | Idle. | on End | “Voice chat ended” |
On each device
- phone
- The main place for voice. The text box becomes a voice bar with status, mute and End, near the thumb. Words appear in the thread above.
- tablet
- As on the phone. A tablet is often propped up, so the text must be readable at arm's length.
- desktop
- The same voice in the chat. End or Stop stays in 1 place and works from the keyboard too.
When to use it
- People's hands or eyes are busy - cooking, driving, fixing something.
- People find listening easier than reading - low vision, reading difficulty, tiredness.
- Answers are short and work spoken - steps, directions, quick facts.
- The answer is a table, code or a long list. It does not work spoken. Use Generative UI and structured output instead.
- The place is quiet or shared - an office, a ward at night. Start in text and offer read aloud.
- People want to speak their ask but read the answer. Use Voice input instead.
The parts of the pattern
- AHer wordsWhat the AI heard, shown as she speaks, so a mishearing is caught early.
- BLive wordsThe answer as text, word by word, in time with the voice. The current word is marked.
- CVoice barSays "Listening" or "Speaking" in words, with a moving wave.
- DEndEnds the voice chat. Always visible, always in the same place.
Do and Avoid
Next, knead the dough for 10 minutes.
〰 Speaking · End
Speaking
Next, knead the dough for
- stopped, you spoke
10 minutes, until it is smooth.
Next, knead the dough for 10 minutes. Then cover it and let it rest for 20 minutes…
(still talking over her)
Writing and accessibility
- Write for the ear: short sentences, the action first, 1 number per sentence. "Knead for 10 minutes."
- Say numbers as people say them: "10 minutes", not "10 min".
- Keep spoken answers short, then offer more: "Want the next step?"
- Show status in words - "Listening", "Speaking" - not only a moving shape.
- Show every spoken word as text, in time with the voice. People who are deaf or hard of hearing need it.
- Speech starts only when the person starts it, and can be paused or stopped at once.
- Announce only state changes: "Listening", "Speaking", "Voice chat ended". Never read the transcript (the written words) out through the status region (the part of the page a screen reader announces on its own). It would talk over the voice.
- End and mute are real buttons, at least 44 px, reachable by keyboard and by switch (a 1-button control some people use instead of touch). The words stay after the voice chat ends.
Seen in real products
- ChatGPT voiceWeb and mobile · Nov 2025Since 25 Nov 2025, voice runs inside the chat: a live transcript and visuals such as maps appear in the thread, and you tap "end" to stop. The old full-screen orb is still there as "Separate mode" in settings.TechCrunch, 25 Nov 2025 ↗
- Claude voice modeWeb, desktop and mobile · Oct 2026Hands-free or push-to-talk. People can speak to interrupt, transcripts are saved, and a Stop button sits at the bottom right on web and desktop.Claude Help: Use voice mode ↗
- Gemini LiveAndroid (beta) · Apr 2026A redesign drops the full screen. A waveform bar takes the place of the text box, with a transcript button, camera and screen share, mute, and a keyboard icon to leave.9to5Google, 19 Apr 2026 ↗
Each example comes from the product’s help pages, release notes or teardowns, with the month it was checked. Products change every month - tell us when 1 is out of date.
Check yourself
Seen this pattern done better, or a product that changed? Press Suggest next to any heading, or tell us about the whole page.