Writing voice can be measured
Writing voice is the recurring pattern of lexical, syntactic, tonal, rhetorical and punctuation choices that makes one person's prose recognisable from another's. Some parts are deliberate, such as formality or preferred vocabulary. Others are habitual, such as function-word frequency, sentence-length variation and punctuation cadence. Stylometry measures these patterns. ScriptGrain applies those measurements across 45 reference points to build a reusable voice profile.
Voice, tone and style are not the same thing
Voice is what stays constant. Tone is what shifts with the occasion: the same writer is warmer in a thank-you note than in a refund policy, and still recognisably themselves in both. Style is the surface layer of conventions, such as whether you write "per cent" or "%".
The distinction matters because tone is easy to instruct and voice is not. Telling a model to be "friendly but professional" changes the tone of the output. It does not make the output yours.
Which parts can actually be counted
A useful test: could two people reading the same page independently arrive at the same number? Sentence length can. "Punchy" cannot.
The measurable parts are not the ones writers tend to talk about. Nobody chooses their ratio of "the" to "a", or how much their sentence length varies from one paragraph to the next. That is exactly why those features identify a writer: they are too small and too frequent to fake deliberately, and they survive a change of subject.
The layers of a voice
A voice profile is not one number. It is a set of measurements across several independent layers, each of which can match or miss on its own.
- Lexical: Vocabulary range, distinctive recurring words, how often you contract, how often you reach for a rare word. Type-token ratio
- Syntactic: Average sentence length and, more tellingly, how much that length varies. Clause order, and whether you build to the point or lead with it. Sentence-length variance
- Tonal: Formality, humour register, and how far you commit versus hedge. Voice vs tone vs style
- Rhetorical: How you open, how you close, whether you argue from evidence to conclusion or the reverse, and how often you reach for a metaphor. Idiolect
- Punctuation: Comma density, dashes, semicolons, parentheses, ellipses. Habits so ingrained most writers cannot describe their own. Markers of AI writing
- Function words: The frequency of words carrying no topic at all: the, a, of, but, however. The single most reliable authorship signal, and the one nobody controls. Function words
- Rhythm and cadence: Whether prose runs punchy or flowing, and where the emphatic sentence lands in a paragraph. Burstiness
- Structure: Paragraph length, list habits, and the recurring shapes a writer builds arguments out of. Stylometric fingerprint
Where the measurements come from
Stylometry is the statistical study of writing style, and it is about a century older than any AI writing tool. It was built to answer authorship questions: who wrote the disputed Federalist Papers, whether two documents share an author, whether a text is consistent with a known writer's habits.
The techniques transfer directly. The same function-word frequencies and sentence-structure statistics that attribute authorship can describe a voice well enough to write toward it, and, more usefully, well enough to check whether a draft actually landed.
What the research shows · What stylometric analysis is · Stylometric fingerprint · Idiolect
Measured voice versus described voice
Most tools ask you to describe yourself. Tone sliders, three adjectives, a paragraph of custom instructions. That is a self-report, and self-reports are unreliable about exactly the features that make writing recognisable, because those features are unconscious.
A measured voice is read from writing you have already done. The difference shows up in what each approach can prove: a description cannot tell you how close a draft came, whereas a measurement can be scored against the profile and given a number.
One practical consequence. Measurement needs material. If you have very little existing writing, a description-based tool will serve you better until you have written more.
What a voice profile is
A voice profile is a reusable record of those measurements, built once from your samples and applied to every later draft. It is not a prompt. It is a target the output can be compared against, which is what makes the result checkable rather than merely plausible.
How ScriptGrain builds a profile
Checking whether a draft actually matches
Generating in a voice and verifying the result are separate problems, and the second one is where most tools stop. Voice Match, ScriptGrain's measure of how closely a draft aligns with a profile, scores it by comparing the measured features directly, so the feedback names what diverged rather than offering a general impression.
This is the part worth insisting on when evaluating any tool in this category, ours included: if it cannot tell you how close a draft came, it is describing, not measuring.
Voice Match, defined · How ScriptGrain scores a draft
Why unguided AI writing converges
Left to itself, a language model writes toward an average: even sentence lengths, low variance, a recognisable set of connectives and hedges. That average is competent and completely anonymous, which is why so much AI text reads as the same voice regardless of who prompted it.
The markers that give it away are the same ones that identify a human writer, measured the other way round.
Markers of AI writing · Burstiness
Measure your own writing
The fastest way to understand any of this is to run it on something you wrote. Both tools below are free and need no account.
Analyse your writing style · Check text for AI markers · Every measured term, defined
Common questions
Can writing voice really be measured, or is that marketing?
The measurable parts are genuinely measurable: sentence-length variance, function-word frequencies, punctuation rates and vocabulary diversity are counts, and two people counting the same text get the same answer. What cannot be reduced to a number is judgement, such as whether an argument is any good. A voice profile describes how you write, not whether it is worth reading.
How is a measured voice different from telling ChatGPT to write like me?
Custom instructions capture what you can describe about yourself, which is a handful of surface preferences. They cannot capture your function-word distribution or your sentence-length variance, because you do not know them. A measured profile is read from your existing writing, and, unlike an instruction, it can be used to score a draft afterwards.
How much writing is needed to measure a voice?
Enough for the frequencies to be stable rather than accidental. One piece is enough to start, and roughly 3,000 words across several samples gives a noticeably firmer profile. Below a few hundred words, measurement has little to work from and a description-based tool will do better.
Does matching a voice mean copying phrases?
No, and the distinction matters. The features that make prose recognisable are structural: rhythm, connectives, punctuation habits, how sentences vary in length. Reproducing those is what makes writing sound like you. Reproducing your actual sentences would just be quotation.
Updated 2026-09-01