How does AI learn your tone of voice from recordings?

23.5%

average relative improvement across seven personalisation tasks, three of them classification, when a fine-tuned model was given items from the person's own history, in the LaMP study

Is the AI really learning from your recordings?

Not in the way most people picture it. A large language model finishes its training long before you open the tool, and a writing tool doesn't need to retrain it for each person. What changes from one founder to the next is the material the model reads just before it writes, which is called its context.

Tone of voice is also a different thing from the sound of your voice. Copying how someone sounds is voice cloning, a separate technology. For a written post, your tone lives in the words you choose, how long your sentences run, the opinions you hold and the stories you keep coming back to.

What happens between pressing record and getting a draft?

1. Transcription. Speech recognition turns the audio into text. This is the step that captures your phrasing, and it is also where names and specialist terms can go wrong.
2. Context. The transcript goes to the model with instructions and, often, a few posts you have written before, so it can see what your finished writing looks like.
3. A brief or memory. Facts you have shared before, such as your company, your audience, the topics you return to and words you never use, are stored and added each time, so you don't repeat them.
4. Drafting. A good tool tells the model to cut, order and tighten, lifting phrases straight from what you said rather than writing fresh ones.

Why does your own material matter more than style rules?

Because the model can only draw on the parts of you it has been given. In the LaMP study, researchers tested what happens when a model is handed items from a person's own history first. A fine-tuned model improved by 23.5 percent on average across seven tasks, three of which sort text rather than write it, with the largest gain on email subject lines.

The same study found that even random items from someone's history usually helped, but the most relevant ones helped more. In our view, a transcript of you talking about the exact topic of the post is about as relevant as material gets.

Style samples alone go less far. A 2025 study of more than 400 authors gave leading models five samples of each person's writing and asked them to copy it. The models did reasonably well on news and email but struggled with informal writing such as blogs and forum posts.

What helped most was the author's own words. Starting a draft from a short snippet of the author's actual text gave the biggest boost in how human the output seemed, with AI detectors judging more of it as written by a person. That is why your own spoken words for a post, plus a growing brief, give the model far more of you to work with.

What should you check in the draft?

Read it once for accuracy. Check every name, figure and technical term against what you meant, because speech recognition still trips over them. Cut anything the model added that you never said, however neat it sounds.

Then read it aloud. If a line doesn't sound like something you would say to a customer, swap in your own words from the transcript. Our guide to turning a voice note into a post walks through the editing step by step.