Speak to ChatGPT and review what it understood
You talk through a quick brief while walking between meetings. Fifteen and fifty can sound uncomfortably alike, so check what ChatGPT heard before it turns your note into a confident plan.

What you'll make
a short checked brief from a spoken note, then a text result built only from that brief.
What you'll need
ChatGPT in your work account with microphone access allowed, a quiet spot, a short note whose facts you can check against something written, and about ten minutes. You don’t need to record anyone else or connect any apps.
Example result.
| Item | Checked detail |
|---|---|
| Trays | 15 |
| Location | Room Linden |
| Contents | Small teal samples |
| Date | Not confirmed |
| Delivery | Not booked |
1. Choose a note you can check
Pick a short note whose facts you know, and keep a written reference for the names, quantities and dates. Say exact dates rather than “next Friday” when the date matters.
Decide what you want at the end, such as a packing list, a short briefing or a tidy paragraph, but don’t ask for it yet. First establish what was said; then ask for the result.
Why this works: if transcription and rewriting happen in one step, you can’t tell whether an error came from hearing or from writing.
What you should see: a note of a few sentences and a written list of the facts that matter.
2. Dictate, or start a voice conversation
These are two different controls:
- Dictation turns your speech into text in the message box, so you can read and edit it before sending. In the ChatGPT desktop app, select Dictate in the composer. On the web and mobile, use the microphone icon in the message box.
- ChatGPT Voice is a spoken back-and-forth. In the desktop app, select Start new voice chat and Stop voice chat when you finish. On the web and mobile, select the separate Voice icon (a waveform) at the right of the message box.
For a short brief, dictation is the better choice: you see the words before anything happens. Read the transcription and check numbers digit by digit, names against your reference, and small words such as “not”, “proposed” and “optional”. A dropped “not” changes the task more than a spelling mistake.
Why this works: dictation puts a review point before sending; Voice is better when you want to think aloud.
What you should see: your note as editable text, or a voice reply you can read in the chat.
3. Ask for a read-back before the result
Whether you dictated or spoke, ask ChatGPT to show what it understood before it writes anything:
Before creating the result, list the facts you understood from my note: names, quantities, dates, locations and anything uncertain. Do not fill in missing details or take any action. I will check and correct this list first.
You can say this in a voice chat or paste it after dictating.
Why this works: a list of facts is quicker to check than a finished paragraph, and errors are harder to hide in it.
What you should see: a short list you can compare line by line with your written reference.
4. Ask for the output from the checked list
Once the list matches your reference, ask for the result:
Using only the checked list above, create [output] for [purpose]. Keep quantities, names and unresolved details exactly as listed. Separate confirmed facts from open questions. Do not send, book or buy anything.
[output] is the form you want, such as a packing list, and [purpose] is who it’s for or what it’s for.
Why this works: “using only the checked list” stops the model going back to the raw transcript, or its own guesses, for detail.
What you should see: the output with every value from the checked list and the open items still marked open.
Check the result
- Each quantity matches your written reference, digit by digit.
- Names and places are spelled as in your reference.
- Anything you said was unconfirmed is still unconfirmed.
- No new action appears (an order, a booking, a message).
The usual failure: a misheard number or a filled-in date. “Fifteen” becomes 50, or “the date isn’t confirmed” becomes “Friday”. Spot it by comparing the read-back with your reference before asking for the output. Name the exact value; “that number is wrong” doesn’t help when there are several:
Correct the list: [wrong value] should be [correct value]. Keep [other checked facts] unchanged. [Detail] is still unconfirmed. Show the corrected list before creating any output. Do not add a booking, purchase, message or other action.
[wrong value] and [correct value] are the misheard and intended values, [other checked facts] are the lines already correct, and [Detail] is the item that stays open. Then read the final output against the corrected list, not your memory of the conversation. A correct list can still become a wrong summary.
Worked example: fifteen trays, not fifty
Fictional example. The spoken note:
For the display practice, we need fifteen trays in Room Linden. The date is not confirmed. Put the small teal samples in the trays; no delivery has been booked.
The read-back said: “50 trays, Room Linden, small teal samples, for Friday.” A later output said “Order 50 trays for Friday.”
- The failure. Three things changed: the quantity (15 became 50), a date appeared (Friday), and an action appeared (order). Only the first is a hearing error; the other two are the model tidying up an open question.
- The repair. The correction prompt with “50 should be 15”, keeping Room Linden and the teal samples, and “the date is still unconfirmed” produced the checked list in the example result.
- Why the read-back mattered. The error was caught in a five-line list, before it was buried in a packing list with headings.
Try it
Fifteen and fifty are different jobs
Switch between the weak rewrite and the checked correction. Highlighted words carry the decision.
Source wording
Fifteen trays, Room Linden, small teal samples. Date unconfirmed; no delivery booked.
Weak rewrite
Order fifty trays for Friday.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Correct things during a voice conversation
In a voice chat you can type as well as speak, and the conversation appears in the chat as text. That helps with numbers: type “15, one five” rather than saying it again. When you end the voice chat, read the transcript in the chat before relying on anything said aloud. What you heard and what was transcribed can differ.
Keep actions separate from discussion
A voice conversation can talk about an action without doing it. In the desktop app, ChatGPT Voice can start and steer tasks, and it follows the same permissions as those tasks. Saying “yes” aloud doesn’t replace an on-screen approval. Check the actual result and destination before trusting that anything was done.
For meeting audio involving other people, read the consent and retention guide before recording.