Feature
Voice Cloning
Upload writing samples. Inksong learns the fingerprint that makes your prose yours, then applies it to the rewrite.
Voice cloning is the feature most paraphrasers quietly skip. They run text through a shuffle of synonyms and sentence flips and hand you back something fluent but anonymous. The cadence is flattened. The verbal tics that mark a piece of writing as yours are sanded away.
We wrote about this in our engineering notes: a humanizer that strips voice produces text that reads “edited by committee.” Inksong takes the opposite approach. You upload one or more samples of your own writing, and the system extracts a quantitative profile of your style before the rewrite begins.
How it works
What's happening under the hood
The voice profile is built along five measurable axes. The first is n-gram fingerprints— particularly the 2- and 3-word sequences you reach for unconsciously. If you start paragraphs with “It turns out” or lean on “in practice” as a pivot, those tics get logged. They will appear in the rewrite at roughly the rate they appeared in your samples.
The second axis is sentence-length distribution. We don't just measure average length; we measure the shape of the distribution — how often you run a long sentence next to a three-word fragment. The third axis, burstiness, is the variance in that distribution. High-burstiness writers swing wildly between short and long sentences; low-burstiness writers hold a steady cadence. Most AI prose is uniformly mid-length, which is what makes it sound metronomic.
The fourth axis is vocabulary specificity— the lexical band you tend to operate in. Some writers stay anglo-saxon and concrete; others reach for latinate abstractions. The fifth is hedging density: how often you qualify with “may,” “tends to,” “arguably.” Academics hedge heavily; journalists hedge sparingly.
These five measurements become a structured style guide that we pass into the Claude prompt — not as literal sample text to imitate, but as a quantitative constraint. We recommend at least 1,500 words of sample text for a stable profile; shorter samples produce noisier fingerprints. Once the profile is built it's attached to your account and applied to every subsequent humanization.
Example
See it in action
It is important to note that the implementation of machine learning models requires careful consideration of several key factors. These include data quality, model architecture, and the selection of appropriate hyperparameters. Furthermore, practitioners must remain cognizant of potential biases that may emerge during the training process and take proactive steps to mitigate them effectively.
Look, getting an ML model to work isn't really about the architecture — it's about the data, which everyone knows but nobody wants to admit. You also pick your hyperparameters, sure, and you try not to bake in obvious bias. It turns out most of the actual work happens before training even starts. The model is the easy part.
Benefits
Why this matters
Consistency across long documents
A 10,000-word thesis stays in one voice end to end. No drift between chapters, no committee-edited middle section.
Recognizable individuality
Your tics survive the rewrite. The people who read your work for a living will still recognize it as yours.
Faster editing afterward
Less manual cleanup. When the output already sounds like you, you spend minutes polishing instead of hours rewriting.
Related features
Start humanizing today
5 documents free a month, no card needed. Three minutes to your first humanized doc.
- 5 documents/month on the free tier
- No credit card required
- Cancel or upgrade anytime