Methodology
Every draft Timbre generates gets a 0–100 voice fidelity score that quantifies how much it sounds like you. Here is the math underneath, in plain English.
A single number is too blunt. We compose the overall score from four layers — each measuring something different about whether the draft sounds like you.
Deterministic, byte-for-byte measurements of how the draft is built — contraction rate, sentence-length variance, paragraph cadence, rhetorical density, vocabulary overlap, voice imperfections, and the function-word stylometric fingerprint. Seven sub-scores share this layer equally.
Cosine similarity between an embedding of the draft and the centroid of your writing samples. Captures stylistic gestalt that mechanical metrics miss — tone, register, characteristic moves at the paragraph level.
A separate model rates the draft against your samples on six voice dimensions (tone, humor, directness, storytelling, hook, closing). Tone-alignment in the popover is the tone slice of this layer; the other five contribute to overall but aren't surfaced as rows.
Dynamic time warping over your paragraph-energy curve. Detects whether the draft rises and falls in the same shape as your samples — the rhythm a reader feels but can't easily name.
The editor popover shows nine numbered rows. Here is what each one measures.
Contractions
How closely your contraction rate (don't, we're) matches the rate observed across your writing samples.
Sentence rhythm
Whether sentence-length variability matches the rhythm pattern of your samples.
Paragraph rhythm
Whether short / long paragraph mix matches your typical paragraph cadence.
Rhetorical density
How often rhetorical questions and devices appear, compared with your sample baseline.
Vocabulary
Presence of your signature phrases and absence of words you tend to avoid.
Voice imperfections
Match on the small irregularities — fragments, dashes, conjunction starts — that make writing feel human.
Function words
Stylometric fingerprint match on the small connective words (the, of, but, so) most predictive of authorship.
Style similarity
Cosine similarity between an embedding of the draft and the centroid of your writing samples.
Tone alignment
LLM judgment of whether the draft hits your tone — formality, warmth, directness — compared with your samples.
Meet Maya — a hypothetical Timbre user whose voice profile is built from her LinkedIn writing. Her samples are casual, contraction-heavy, fragment-friendly, and short.
“Most onboarding flows are bloated. We cut ours from twelve steps to four. Conversion didn’t just inch up — it doubled. Turns out, less to fill out means more people actually finishing. Who’d have guessed.”
Draft #1 — overall fidelity 52
“We have observed that streamlining the onboarding process from twelve steps to four resulted in a substantial increase in conversion, effectively doubling the previous rate.”
Draft #2 — overall fidelity 81
“Onboarding was bloated. We cut it from twelve steps to four — and conversion doubled. Turns out, less to fill out means more people finish.”
Same idea. Same facts. Two very different scores — because fidelity measures how the words land, not whatthey say. When you edit a draft, Timbre learns which way you push the score and biases the next generation in that direction. That’s the loop.
Connect a few samples, generate one post, and the popover lights up. Edits start teaching it within a session.
Try it on your own writing