02
Story
When my hands are busy or typing is slow, I want to speak my ask and see the words before they are sent, so a misheard word doesn't give me a wrong answer.
Joseph, a maize farmer near Eldoret, Kenya, asking about fertiliser on his phone with soil on his hands.
- 01 Setting
Joseph is in his field and wants to know how much fertiliser to buy.
- 02 Trigger
He taps the mic and says "2 acres". It hears "10".
- 03 Without it
The message goes as soon as he stops. The answer is for 10 acres, and he nearly buys 5 times too much.
- 04 With it
The words land in the box. "10" is marked as unsure. He taps it and picks "2".
- 05 Outcome
He sends, and gets the amounts for his 2 acres.
03
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.
After speaking
Mic
Device
04
What happens, step by step
| State | The person sees | The AI is | When | Screen reader |
|---|---|---|---|---|
| 1 | A mic in the box, with a hint: "Tap the mic and speak". | Idle. | - | - |
| 2 | A pulsing dot, a timer, a level meter and the words as they are heard. Stop and Cancel are always visible. | Turns speech into text. | while speaking | “Listening” |
| 3 | The words sit in the box, not sent. Unsure words are marked. | Idle. | on stop | “Stopped listening. Check the text, then send” |
| 4 | Tapping a marked word offers other words it might be. 1 tap fixes it. | Idle. | on tap | - |
| 5 | The avoid version: the message goes as soon as listening stops, mistakes and all. | Reads the misheard ask. | on stop | “Message sent” |
| 6 | The sent message is marked "Spoken". The answer streams in (arrives word by word). | Writes. | - | “Generating a response” |
| 7 | An answer for what he meant. | Idle. | when done | “Response complete” |
05
On each device
- phone
- The mic sits next to Send, by the thumb. Hold to talk suits noisy places. Tap to start and stop suits longer asks.
- tablet
- As on the phone. In split view, the listening bar stays inside the app's own box.
- desktop
- The mic sits in the box with a keyboard shortcut. Typing while listening stops listening and keeps the words.
06
When to use it
Use when
- People's hands or eyes are busy - cooking, driving, working outside.
- Typing is slow or hard for people - on a phone, in a second language, or with a motor disability.
- Asks are long and spoken words come more easily than typed ones.
Don’t use when
- People want a spoken conversation, with the AI talking back - that is voice mode. Use Voice output and live transcript instead.
- People need to show something as well as say it. Use Multimodal input instead.
- The ask is a short choice from a list - taps are faster than speech. Use Quick actions instead.
07
The parts of the pattern
- AStopThe mic turns into a stop button, with a label: "Tap to stop" or "Release to stop".
- BListening statusA pulsing dot and a timer. People always know the mic is on.
- CLive wordsThe words appear as they are heard, so mistakes show early.
- DCancelThrows the recording away. Nothing is sent or kept.
- ELevel meterMoves with the voice, so people know they are being heard - or not.
08
Do and Avoid
Check the words, then send.
…for 10 acres of maize?
…for 10 acres of maize?
Sent as soon as he stopped.
How much fertiliser do I need…
● 0:03 · Tap to stop · ✕
…
●
09
Writing and accessibility
Writing
- Label the state, not the tech: "Listening", not "Speech recognition active".
- Name how to stop: "Tap to stop", "Release to stop".
- Say what to do next when listening ends: "Check the words, then send."
- Mark spoken messages in the thread: "Spoken".
Accessibility
- Announce "Listening" and "Stopped listening" in the polite status region (the part of the page a screen reader announces on its own). Never read the live words out while the person speaks.
- The mic button's name changes with its state: "Start dictation", "Stop dictation".
- Keep a visible label on every button that matches its name, so people who use voice control can say it.
- Never auto-send on silence. People who speak slowly, stammer or pause to think need time.
10
Seen in real products
- ChatGPTNot stated · Nov 2026Dictation turns the recording into text that people can edit before sending as a message.OpenAI Help: Voice dictation FAQ ↗
- ClaudeiOS and Android · Nov 2026A mic on the right of the box turns speech into text in the chat; people tap the arrow to send or "X" to cancel. 12 languages, most in beta.Claude Help: Use dictation on Claude Mobile ↗
- Claude voice modeWeb, desktop and mobile · Nov 2026Hands-free listens and answers at natural pauses; push-to-talk is for noisy places. Text transcripts are saved in the chat.Claude Help: Use voice mode ↗
- ChatGPT voiceWeb and mobile · Nov 2025Voice runs inside the chat, so people see answers appear while they talk. They tap "end" to stop; the old full screen is a setting.TechCrunch, 25 Nov 2025 ↗
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.
12
Check yourself
Sources
Suggest an edit
Seen this pattern done better, or a product that changed? Press Suggest next to any heading, or tell us about the whole page.