Voiceprint Studio
Voiceprint Studio builds and sharpens an email style guide, the document you paste into any AI so it drafts email in your own writing voice. It grounds each draft in your real sent emails, can self-train against a held-out test set, and turns each miss into an approved rule, across four tabs. Still in development, and it never sends email itself.
On this page
What it is
A first-party Omniscio sidebar panel that builds and sharpens an email style guide — the document you paste into any AI so it drafts email in your voice. "Voiceprint" here means your writing voice, NOT audio / TTS / spoken responses. In-development; gated via the voiceprint-studio unreleased feature (default off — reveal via Settings → Lab toggle or AMC_SHOW_VOICEPRINT_STUDIO=1). Renders as a native built-in virtual project beside the projects sidebar (NON-spawnable — it hosts no Claude session).
Where to find it
It renders as a native built-in project beside the projects sidebar, so you open it like any other panel and it hosts no Claude session of its own. Because the feature is still in development it stays hidden until you reveal it, from the Settings → Lab toggle or the launch switch named in the paragraph above.
How it behaves
What it's for
Capturing a person's email voice well enough that an AI can pass for them takes a calibration loop: show the AI a real email you sent, have it try to write that same email blind, compare, and turn each "that's not how I'd say it" into a rule. Voiceprint Studio is that loop as an app. Two levers make it converge: the AI grounds each draft in the person's REAL sent emails (few-shot), not the guide text alone; and it can self-train — draft a held-out test set, grade itself against the real replies, refine the guide from the worst miss, iterate a few rounds under a budget, and only then surface the hard cases to you.
The four tabs
- Corpus — Pull your Gmail Sent folder: your whole sent history, all years, across every connected account (not just the newest few hundred). The first pull pages through everything and can be interrupted or restarted — it resumes from where it stopped without duplicating emails — and every later pull is incremental, reading only what's new since the last one (a backlog bigger than one round is paged through over several rounds, never cut short). A round pulls a bounded batch and the panel keeps asking while accounts remain pending, and if an account genuinely fails the pull stops, names that account and re-enables Pull instead of retrying forever. Each email records which account it was sent from, so a work voice and a personal voice stay separable. Each email is cleaned to the words you actually wrote: quoted replies (Gmail, Outlook and Apple Mail), forwarded-message blocks,
>lines, signature blocks and legal footers are stripped before anything reads the body. Automated, newsletter, calendar, quote-only, empty and duplicate mail is dropped with a recorded reason, and the rest gets an auto-guess — You (hand-typed, ground truth) or AI (Claude-drafted / skip) — with the plain-English reason shown; confirm or flip with one click. The corpus splits into a training pool (grounds drafts) and a held-out test set (the self-grade eval — never used as an exemplar). AI-suggested chain exclusions surface the drops a single row cannot see (duplicates, templates) for one-click confirmation. Your mail is then split into your separate voices, each its own profile with its own corpus/guide/training. - Sensitive mail — Every email is checked before it can shape your voice. A local check runs on your computer first, and anything it flags — health, legal, money, work/HR, passwords and codes, personal relationships, someone else's private information — is held back: mail that fails the on-device check is never sent to any model. Mail that passes it goes on to be labelled by kind (quick reply, short reply, substantive, scheduling, introduction, request, decline) by the Jev classifier, and anything that comes back sensitive — or unlabelled, because the labeller could not answer — is kept out of training, example drafts and the blind test. Until an email has been checked it is treated as sensitive too, so a fresh or interrupted pull holds mail back rather than letting it through. A Sensitive mail section lists what is held back and why, and you can put one back with one click.
- Interview (Calibrate) — Pick a real You email. The AI drafts it blind and grounded in your real training emails — it sees the prior context + guide + a few of your most similar real sends, never the target's real reply. Side-by-side: the incoming context · what you sent · what the AI wrote, with a 5-point "how close?" rating and the AI's own judge score. Blind mode hides which pane is which (server-fixed order) until Reveal. On a miss, a short clickable interview (the app's QuestionWidget format) turns the gap into an approved guide rule. A self-training panel starts an autonomous run (test-set size / rounds / budget) and shows the score trajectory.
- Style Guide — The living guide: editable, versioned, with history. Approved fixes append here, each traced to the comparison that produced it.
- Progress — Pass rate + the objective average judge/closeness score. When it climbs and stays high, the guide has captured the voice.
Your voices
People do not write in one voice. The same person opens a client email with "Dear Ms. Chen" and closes it "Kind regards", writes a colleague "Hi Sam — thanks for the quick turnaround", and texts a friend "Hey Mike, that was awesome!". Voiceprint Studio works that out for you instead of guessing from the recipient's address book domain.
Find my voices reads the cleaned corpus and groups it by three things it can actually see: the register of the writing (the greeting and sign-off you habitually use, how often you contract and exclaim, your informal and formal vocabulary, how long your sentences run), the account you sent from, and the kind of recipient (a personal address, a company address, or unknown). A clear gap in register inside one account is two voices, not one. Each voice is named plainly — Personal, Work, and a register word such as Work — everyday when two would otherwise share a name.
Nothing leaves your computer to do this: it is a local, rules-based pass over mail you already pulled, with no model call and no cost. Discovery runs automatically when a pull finishes, and you can run it again any time.
Two controls on each voice:
- Rename — your name for the voice, and the discovery never overwrites it, however many times it runs afterwards.
- Merge into… — folds one voice into another: its emails move across and the merged-away voice, with its guide, is removed. The merge is remembered, so the next run does not bring it back.
Emails are never double-counted: each one belongs to exactly one voice, and running discovery twice on the same corpus changes nothing. The two main voices — the one that mostly writes to people and the one that mostly writes to companies — keep the ids the rest of the app already drafts with, so the voice that reaches your drafts keeps reaching them. A third voice trains its own guide but is not yet consulted when a draft is chosen for a recipient.
The honest "sounds like me" score
The Progress tab's self-grade is useful while the loop is running, but it is not a held-out measurement: the same Claude model that drafts also grades, on emails the guide is refined against. The blind voice score answers the question honestly.
A fixed set of your real sent emails is marked held out — training, the guide and every example the drafter is shown are barred from it in the query layer, so it cannot leak into the work being measured. Scoring walks that set: for each email the current guide drafts a blind reply, then a judge running on a different model family than the drafter (GPT-5.6 Luna through OpenRouter, always that exact model — a Utility AI model you saved in Settings never replaces it) is shown your real reply and the AI draft in randomised order and asked which one you wrote. Because the sides are shuffled, a perfect imitation is a coin flip: 50 means indistinguishable, 0 means obviously not you, and the same run reports a real-vs-real control that must sit near 50 or the judge itself is suspect. Each run stores its score, the control, the guide version and the exact spend, so a later run reports the before/after delta. A failure that would route the judge back onto Claude's family is detected and the pair discarded rather than counted.
Requirements & graceful states
Pulling the corpus needs Gmail connected in Omniscio; generating a blind draft or a rule needs an Anthropic API key configured. If either is missing the panel shows a plain message ("Add an Anthropic API key in Omniscio to generate drafts", "make sure Gmail is connected") rather than failing. Each blind draft / rule proposal is a small metered Claude call.
Privacy / data
Your sent emails are stored locally in Omniscio's SQLite DB only (your machine). The blind-draft context is built to exclude your real reply (a unit test locks this), so the model is never shown the answer it's being compared against. Erasing a persona hard-deletes its stored mail and also resets the account's pull watermark, so the next pull re-reads your history for the remaining (or a re-created) persona instead of trusting a watermark that would only ever fetch new mail.
Two things leave the machine, both only when you start them: the metered Claude draft calls, and — for the blind voice score only — the pair of email texts sent to a judge on a different model family (GPT-5.6 Luna through OpenRouter). The judge is what makes the score independent, so this is deliberate and owner-approved; the judge never sees anything beyond the two texts and the short situation line both are answering. That situation line comes from your thread, so it is screened by the same rule as the corpus: an earlier message that is sensitive or not yet checked is dropped before either model sees it, and the held-out set the score walks is drawn only from checked-clean mail.
The Learn-my-voice chat — the main road
Voiceprint Studio's setup now starts in one chat, the first and default tab of the panel. It is the place the owner actually runs setup; the four detailed tabs remain below it as the inspection road.
- Read your Sent mail — one pull of the whole history across every connected account, then a plain sentence: how many sent emails were found and how many were kept as really yours. One click opens the drop reasons, with real example subjects.
- Voices I found — each voice with its email count and three real examples, and Approve / Rename / Merge / Split. This card is not connected to voice discovery yet, so it says "coming soon" rather than showing invented voices. Discovery itself already works: it runs after every pull and from Find my voices on the Corpus tab (see "Your voices" above).
- Your top habits — the guide's rules in plain English, each with Keep, Fix or Remove. Keep writes nothing; Remove and Fix save a new guide version the moment you click.
- Proof — three of your real sent emails, each answered blind, shown beside what you actually sent with the judge's voice-match score. Approving the voice is a separate, explicit click.
- Later learnings — anything new learned about your voice arrives as a one-line approval card; nothing changes the guide unless you approve it.
Every step is honest about its limits: the proof card states what really leaves the machine for that one step — three of your sent emails, a few more sent emails as writing examples, and the earlier messages in those threads — and a step whose backend is not built yet says so rather than guessing.
Related
The corpus this panel trains on comes from Gmail, which Gmail Integration covers, and what the finished guide is for is set out in Agent Email. If your interest is drafting with AI rather than capturing a voice, AI Writer is the sibling surface.
Last verified 2026-11-14