voiceprint

Turns your own spoken words into a finished article. AI does the research,
structure, headings and checking. You supply the sentences, by talking.

Why it works

Text reads as human in proportion to how much of your literal wording survives
into it. Not your ideas, your sentences. Measured on one article, varying only
how much was model-composed:

It is a threshold, not a slope. About one sentence in eight can be written for
you. That covers joins and transitions, not sections.

Install

git clone https://github.com/charlie947/voiceprint.git
cd voiceprint
bash setup.sh

That creates your folders, checks what you have, installs the Claude Code skill
if you want it, and runs the tools against a bundled example so you can see they
work.

Then drop your transcripts into voices/yourname/ and type /voiceprint.

Install, the long way

Claude Code: copy the voiceprint folder into ~/.claude/skills/, then
type /voiceprint.

Claude, ChatGPT or Gemini in a browser: open PROMPT.md and follow it. It
is written to be pasted straight into a chat box, in three short messages. No
install, no Python.

Use

1. Talk for 11 to 13 minutes per 1,500 words you want. Record it, or dictate
   into Google Docs with Tools > Voice typing, which is free with no length cap.
2. Get a raw transcript. Do not tidy it.
3. Run the skill. It plans structure, then cuts your transcript down.
4. Check it. On Claude Code, save your transcript as transcript.txt and the
   draft as draft.md in the same folder, then run
   python3 trace.py draft.md transcript.txt. It fails below 85%. In a browser
   chat, PROMPT.md has a manual version of the same check.
5. Retype any sentence it flags, in your own words.

See EXAMPLE.md for a real before and after, with the numbers.

Try it in 30 seconds

The example/ folder holds a real transcript and the article cut from it.

git clone https://github.com/charlie947/voiceprint.git
cd voiceprint
python3 trace.py example/article.md example/transcript.txt

That article scored 100% human. The same recording written up by a model
instead of cut down scored 100% AI.

What this is not

Not a humaniser and not a detection bypass. Style rewriting, humaniser prompts
and register-matching were all tested and all made results worse. This produces
writing that genuinely is yours, because you said it.

If you will not record yourself, this does nothing for you.

Honest limits

It predicts a verdict, it does not guarantee one. Genuinely human writing is
misclassified sometimes, and detectors misfire hardest on non-native English
speakers, autistic writers and formulaic genres. Keep your recording, raw
transcript and draft history — provenance beats any score.