The three steps
1. TALK. Open the voice recorder on your phone and say the thing out loud, the way you would explain it to a client sitting in front of you. Two minutes is plenty. Do not write anything first, that defeats the whole point.
2. TRANSCRIBE. Get the text of exactly what you said. Every "what I've found is", every repetition. The mess is the useful part, because the mess is how you actually talk.
3. CUT. Now delete. Remove the false starts and the repeated line you said three times. What is left is yours and it already sounds like you, because it is a recording of you.
The rule that makes it work: cut, do not rewrite. The moment you start improving sentences you are back to writing, and it stops sounding like you again.
How to check it honestly
There is a free open-source tool called voiceprint that scores what percentage of a finished draft traces back to a real transcript of you speaking. It is by Charlie (charlie947 on GitHub), MIT licensed, and it runs locally.
What it showed me: a post I wrote myself, carefully, in what I thought was my own voice, scored zero percent. Not one sentence traced back to how I actually speak. The same post rebuilt from a recording of me saying it scored 88.8%. Same idea, same person. The only difference was whether I had said it out loud first.
On AI detectors: if the goal is to pass a detector, you have already accepted the writing is not yours. The detector is not the audience. The person reading it is, and they can tell long before any tool can.