Speech-to-email: transcript in, constrained rewrite out, facts checked

Step 1: Talk it out

Use whatever you already have:

• iPhone Voice Memos. On iPhone 12 or later, Apple's Voice Memos app can transcribe what you say, in English and several other languages. Record, then open the recording and choose Copy Transcript from its options.
• The microphone on your keyboard. On iPhone and Android, tap the microphone on the keyboard and speak straight into a notes app or the chat box.
• The chat tool itself. ChatGPT's app lets you dictate into the message box, or talk to it out loud like a phone call.

Start by saying who it's to and why: "Email to Ruth at the bookkeeper. Three things about September." Then just talk. Spell any unusual names.

Step 2: Ask for a clean email

Paste the words and send:

Turn this voice note into a short, friendly email to
[NAME]. Keep every fact, date, and number exactly as I
said it. Don't add anything I didn't say. Use short
paragraphs. If something is unclear, put it in [brackets]
instead of guessing.

"Don't add anything" is the important line. Without it, AI likes to add polite promises you never made, like "I'll follow up next week."

Step 3: Check the facts that matter

Transcripts mishear things. "Fifteenth" becomes "fiftieth." "Ruth" becomes "roof." Before you send, check:

• Every name, including the person it's going to.
• Every number, price, and date.
• Anything in [brackets].
• Anything that commits you to something.

Fix them in the email, not by asking the AI again. It's faster.

Make it sound like you

If it comes out stiff, ask: "Make it sound more like me: casual, warm, no corporate words." Or paste one of your old emails and say "Match this tone." Then read it out loud once. If you wouldn't say a line, change it.

What not to say into it

If the email is about a customer's private details, health, or money, leave those out of the recording, or swap them for placeholders before you paste. You can put the real details back in yourself.

Under the hood

Two error sources stack here:

• Speech recognition errors hit proper nouns and numbers hardest: names, ordinals, prices. They look plausible in the transcript, so they survive a quick skim.
• Generation drift adds content the speaker never said. Polite closers and soft commitments ("I'll follow up Friday") are the typical additions.

The prompt handles the second: "keep every fact," "add nothing," and "bracket what's unclear" turn an open rewrite into a constrained one. Only you can handle the first, because the model can't hear the original audio. A quick entity pass (names, numbers, dates, commitments) catches what's left. For recurring emails, save the prompt with a tone sample so output stays consistent.

Short close

Say it out loud, copy the transcript, and ask for a short email that keeps every fact and adds nothing. Check names, numbers, and dates, then send.