Why we built a voice cloner instead of just a paraphraser

title: "Why we built a voice cloner instead of just a paraphraser" description: "Paraphrasers homogenize prose. A good humanizer should sound like the writer who used it. Here's how Inksong's Voice Cloning works under the hood." date: "2026-05-11" tag: "Engineering" author: "Inksong"

The first time we ran an AI-generated paragraph through a generic paraphraser, the output was technically different from the input. It was also indistinguishable from any other paraphraser's output. That sameness is the point we kept coming back to: tools that "humanize" text without a model of the human end up producing prose that sounds like the paraphraser, not the writer.

So we did the simpler thing first — and shipped the harder thing next.

What a paraphraser is doing

Most off-the-shelf paraphrasers run a single pass: swap synonyms, reorder clauses, lower-cohesion the sentence-to-sentence flow. The output is grammatical, the meaning survives, and the AI-detection score drops a little. But the fingerprint of the writer is gone. A pastor's sermon comes out reading like a marketing blog. An academic abstract comes out sounding like a college essay.

This is fine for the use case of "make this sentence not sound like ChatGPT." It's not fine for the use case of "make this draft sound like me."

What Voice Cloning actually does

When you upload a writing sample to Inksong, we extract a profile across several axes:

N-gram fingerprints. Which 2-and-3-word sequences appear unusually often in your writing? "It turns out" might be a personal tic; "It is important to note that" probably isn't yours.
Sentence-length distribution. Some writers default to a 12-word rhythm. Others swing between 4-word punches and 28-word digressions. We model the distribution, not just the average.
Burstiness. Closely related: how variable is your sentence length, paragraph to paragraph?
Vocabulary specificity. Do you reach for "however" or "but"? "Examines" or "looks at"? We capture the lexical band you write in.
Hedging density. Academics hedge more than marketers. Even within academia, you hedge at a specific rate.

The profile becomes part of the prompt we send to Claude for humanization. Not as a literal style sheet — that produces brittle, robotic output — but as a soft guide: write so the output looks like it came from a writer whose tics are X, Y, and Z. Claude is genuinely good at this.

Why this is harder than it sounds

The temptation is to feed Claude a few sample paragraphs and say "write like this." That works for a paragraph or two, then collapses. By the third paragraph the model has drifted to its own neutral house style. So our prompt isn't "imitate this sample"; it's "match these specific extracted features." The features survive longer than vibes do.

The other temptation is to over-fit. If your profile is built from a single 200-word sample, the model latches onto incidental quirks — a specific phrase you happened to use, a word choice that wasn't really yours. We weight features by frequency and require a minimum sample length before a profile is usable.

What we deliberately didn't build

We didn't build a writing assistant. Inksong takes finished AI-generated text and humanizes it. It doesn't autocomplete, doesn't suggest, doesn't catch typos. Grammarly does that well; we don't need to.
We didn't build a paraphraser. Inksong rewrites whole documents with a coherent voice held across paragraphs. Sentence-level paraphrase is a side effect, not the product.
We didn't build an AI detector. We do show a heuristic before-and-after score, but we're upfront that it's a heuristic. Read the post about that here.

Where this is going

The next step is per-character voice profiles for fiction writers — different voice for dialogue from different characters, all from a single chapter sample. That's on the roadmap, not in the product yet. If that sounds useful, tell us.

If you want to try the current Voice Cloning, sign up free — you can upload a sample, train a profile, and humanize a document inside of five minutes.