The Voice Fingerprint: What It Takes for AI to Sound Like You
By Jack Stovell · 2026-08-13 · Opinion
The Voice Fingerprint: What It Takes for AI to Sound Like You
Building ScriptGrain has forced me to think about voice more carefully than I ever did as a writer. I don't mean brand voice in the mood-board sense, all adjectives and aspiration. I mean the measurable stuff: the habits in your writing that make a paragraph recognisably yours before anyone checks the byline.
I've started calling that a voice fingerprint. This post covers what one is made of, what building profiles for wildly different styles taught me, and how to start capturing your own.
What a voice fingerprint is made of
When ScriptGrain analyses a set of writing samples, it extracts a stylometric profile across 45 measured attributes. Some are mechanical: sentence-length spread, punctuation habits, how often paragraphs open with a conjunction. Others sit closer to personality, like how much you hedge, how directly you address the reader, and how formal you stay when you're explaining something difficult.
No single attribute is impressive. Everyone uses commas. The combination is the fingerprint, and it's surprisingly hard to fake, which is exactly what makes it useful. Hand that profile to an AI model and it stops producing the smooth, interchangeable prose every assistant defaults to. It starts producing something with your grain in it.
My test for whether it's working is simple. Generate a paragraph, put it next to one you genuinely wrote, and ask someone who knows your writing to spot the fake. When they hesitate, the profile has caught something real.
What extreme profiles taught me
Early on I built profiles from deliberately extreme sources, because extremes show you whether the analysis measures anything real. One came from a Victorian novelist, all winding clauses and theatrical flourish. Another was a stripped-back operator register: short sentences, no ornament, straight to the point. A third captured a bestselling self-help author's punchy, mildly confrontational style, and a fourth came from a UK challenger bank's public copy, that warm-but-efficient tone a bank shouldn't be able to pull off but somehow does.
The Victorian one still makes me laugh. Ask it for a product update and the result reads like I'm narrating a coal mine disaster. But the exercise proved the point. The same brief, run through different fingerprints, produces genuinely different writing: the rhythm and the sentence shapes change, and so does the attitude towards the reader. That's a deeper change than swapping a few adjectives on the same skeleton.
Two lessons from that phase carried over into everything since. Range beats volume: five samples from different contexts teach the system more than twenty versions of the same newsletter. And strong voices survive the process better than bland ones, so if your writing has quirks, feed the system the quirks rather than the polished version you think a professional should sound like.
Phrasing is the easy half
I test all of this on the harshest available user: me. On the side I've been building a personal voice assistant that drafts notes and messages in my own voice, and it exposed the limit of pure phrasing faster than any customer could have.
Matching how I phrase things got convincing quickly. Matching how I think did not. Anyone can mimic surface phrasing with enough samples. The harder layer is the pattern underneath: the logic you default to, the questions you ask before committing to anything, where your thinking jumps before it lands somewhere useful.
That matters to you because it marks the honest ceiling of any voice tool, including mine. A fingerprint built from your writing captures how you sound. It can't know what you would choose to say, or which half-formed idea you'd kill on sight. So the workflow that works keeps you in the judgement seat: the AI drafts in your voice, you decide what survives. Be suspicious of anything that promises to remove that step.
How to capture your own
If you want to try this, with ScriptGrain or with careful prompting on your own, the process is the same.
Start by gathering five to ten samples that genuinely sound like you. Emails you didn't overthink beat polished blog posts, because polish sands the fingerprint off. Then make sure they cover more than one context; a single context teaches a costume, several teach a voice.
Next, interrogate what comes back. A good profile should tell you things about your writing you half-knew but never articulated, like how often you hedge (in my case: constantly, then I delete most of them in the edit). If it only tells you things you'd put in a bio, it hasn't caught anything.
Finally, test side by side and iterate. Generate something, compare it against the real thing, and note precisely where it jars. Feed that back in. The first pass is rarely right; the third pass is often uncanny.
The quickest way to see your own fingerprint is to run a few samples through the ScriptGrain voice preview and read what it picks up. Mine is still smudged in places. It's getting clearer.
---
By Jack Stovell, founder of ScriptGrain. I build the product this blog writes about; everything here comes from shipping it. See how ScriptGrain works.