voiceprint
github.com/charlie947/voiceprint- Category
- Productivity
- Rank
- No. 2994Tools index
Previous survey · No. 1935 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Productivity & Collaboration
- Interfaces
- CLI
- Builder
- charlie947
- GitHub
- 67 stars
- Date
About
Voiceprint converts a raw spoken transcript into a finished article by having AI handle research, structure, and headings while preserving the user's own literal sentences rather than paraphrasing them. It includes a scoring script that predicts how 'human' the result reads based on what percentage of original wording survives, requiring roughly 85% verbatim retention to pass.
What it does
A writing method packaged three ways: a Claude Code skill, a paste-in prompt for any chat assistant, and a small set of Python checkers. You talk through a topic out loud, then the model plans the piece and edits your recording down by deletion only. A matching tool compares each line of the finished draft against what you said and lists anything the model made up, so you can say it again yourself.
Why it's ranked here
The idea is narrow and the discipline is real. The instructions forbid the model from rewriting, tidying the raw text, or answering its own gap questions, and the checker exits with a failure code when too little of the draft traces back to the recording. The whole thing is MIT licensed and the core checker needs only the Python standard library. The weak point is the evidence: the threshold it enforces comes from one article tested against one detector, which the author states openly.
What's good
The rules are specific enough to follow. Allowed edits are listed (cut, reorder, join two fragments, fix misheard names, write headings) and so are forbidden ones. The browser prompt splits setup, raw input and drafting into separate messages so the model reads the rules before it sees your words, and includes a recovery message for when it drifts back into rewriting. Setup asks before writing anything outside the project folder, and local transcription keeps audio on your machine.
Tradeoffs
Several quick-start commands in the readme point at locations that do not match the repository layout, so the thirty-second demo fails as written until you fix the paths. The readme says the browser prompt takes three messages while the prompt file says two and then lists four. Local transcription runs only on Apple Silicon. A second scoring tool is hardwired to the author's own voice data and refers to a companion script and folder that are not in the repository. There is no test directory.
How to use it well
Budget about eleven to thirteen minutes of talking for each fifteen hundred words you want, and keep the raw text untouched. Use the plan step to collect missing questions, then answer them out loud rather than by typing. Run the checker on every draft and retype the flagged lines yourself instead of asking for better ones. It suits newsletters, blog posts and long social posts from people comfortable recording themselves. It does nothing for anyone who will not talk.
Technical notes+
bin/trace.py splits the draft into sentences (skipping headings, bold-only lines and fragments under three words), strips stopwords, anchors each sentence on its largest difflib match in the transcript, then counts how many of its content words appear in a padded window around that anchor. Word order inside the window is ignored. Sentences are tagged VERBATIM at 0.85 or above, EDITED from 0.55, COMPOSED below; VERBATIM plus EDITED words must reach the default 85 gate, overridable with --gate, and the process exits 1 on failure. bin/profile.py mines openers, fillers and repeated n-grams per speaker and clamps band floors at zero; bin/spoken.py builds bands without that clamp and defaults to a single hardcoded profile. bin/transcribe.sh shells out to mlx-whisper through uvx. setup.sh copies skill/SKILL.md into the user's Claude skills folder only after a y/N prompt.
Observed
- License
- MIT
- Language
- Python and Bash
- Interfaces
- Claude Code skill, paste-in chat prompt, command-line checkers
- Checker dependencies
- Python standard library only (bin/trace.py)
- Local transcription
- Apple Silicon only, via mlx-whisper and uv (bin/transcribe.sh)
- Default pass gate
- 85 traceable, configurable with --gate (bin/trace.py)
- Install
- git clone plus interactive setup.sh
- Tests
- No test directory in the repository tree
Read from README.md, skill/SKILL.md, prompts/PROMPT.md, setup.sh, bin/profile.py, bin/trace.py, bin/spoken.py, bin/transcribe.sh, docs/EXAMPLE.md, LICENSE.
What it can do
Convert a raw spoken transcript into a structured article with AI-generated headings while preserving the speaker's original sentences
Spoken transcript text → Finished article
Score how much of the original wording was retained verbatim in the output
Transcript and generated article → Verbatim retention percentage
Tags
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.