Every other dictation app sends your voice to a server to be transcribed. Lexpert Voice runs the whole thing on your own computer. No cloud transcription. No account required to use it. Nothing about your matters ever goes anywhere you didn't send it yourself.
When you dictate into Lexpert Voice, your microphone audio is turned into text by a language model that lives on your hard drive. That audio, the transcript it produces, your legal dictionary, and every client and matter name you say stay on your machine. They are never transmitted, never stored on our servers, and never used to train anything. There is exactly one thing that ever leaves your computer, we tell you precisely what it is below, and you can watch it go.
Most tools ask you to trust a privacy policy. We show you the wire.
We split the app into two channels and keep a hard wall between them. One channel holds everything about your work and never opens. The other carries a short, boring list you can read in ten seconds.
The confidential channel. This wall does not open, ever.
Only if you choose to create an account.
That is the whole list. No audio. No transcript. No names. No dictionary. No client data. If it touched a matter, it is in Channel B and it is not here.
# POST /v1/usage — the complete payload, nothing hidden { "account": "you@yourfirm.com", "app_version": "1.2.0", "period": "2026-07-05", "words_dictated": 4213, "sessions": 9, "packs_active": ["tx-family-law"], "features_used": { "push_to_talk": 9, "dictionary_edits": 3 }, "crashes": 0 } # No transcript. No audio. No client or matter names. Ever.
Turn off the account and even this stops. The dictation keeps working exactly the same, offline, forever.
The moment your client's words pass through someone else's server, a vendor is holding their confidences. Local processing removes that question before it can be asked.
There is no vendor sitting on a copy of your dictation, no server to breach, no third-party log to be compelled in discovery. What isn't collected can't be produced.
Your duty of confidentiality asks for reasonable efforts against disclosure. The strongest version of that effort is simple: don't disclose it at all. Nothing leaves, so there is nothing to safeguard in transit.
Your transcripts don't become someone's training data. The model was trained before it reached you. It reads your words to type them, then forgets them.
No outbound connection is required to dictate. Fine for firm laptops with the network clamped down, fine on a plane, fine in a courthouse basement with no signal.
The only thing that ever leaves is the short account list above, and we put it on this page instead of burying it in a policy. That is the whole deal.
If you sign in, your login token lives in the macOS keychain, never in a plaintext file. Sign-in is passwordless and only needed once.
Plain dictation types the sounds you make. Siri tries to be an assistant and answer you. Lexpert Voice does one job that neither does well: it works out the word you were reaching for, spells it correctly, and writes down what you settled on. Nothing more.
Plain dictation hears "voir dire" and types "voider." It mangles case names on contact. A scribe with a legal vocabulary recognizes the term and gets the spelling right.
You stop, restart, and correct yourself out loud. It resolves your false starts and "strike that" to the words you actually landed on, the way a good court reporter would.
It recovers the word you meant to say. It never changes what you meant. Ask it a question mid-sentence and it writes the sentence down. It does not answer.
"The court proceeded to voider the panel."
The court proceeded to voir dire the panel.
✓ Recovered the word you were reaching for. Same meaning, correct spelling.
"Send it to the respondent, no, to the petitioner's counsel."
Send it to the petitioner's counsel.
✓ Resolved your self-correction to the words you settled on. The false start is gone.
"Under obergefel v hodges, same-sex marriage..."
Under Obergefell v. Hodges, same-sex marriage...
✓ Fixed a mangled case name against the pack glossary. Verified citations only, never invented ones.
"The judge threw it out."
The judge threw it out.
The line it will not cross. It will never "improve" this into "the court sustained the objection." That is interpreting what you meant. Not its job.
Smart enough to spell it right. Never smart enough to put words in your mouth.
The word-recovery step runs under hard rules, and you get the final say on every change. It is help you can predict, which is the only kind a lawyer can use.
It corrects spelling and cleans up disfluencies. It does not touch the substance of what you said, the parties you named, or any figure you gave.
Nothing is silently rewritten. You see exactly what changed from what you said, and you accept or reject it.
If your dictation contains a question or a command, it writes those words down like any others. It will not answer the question or run the command. It is not an assistant wearing a scribe's coat.
Accuracy is built up in stages, from cheapest and most certain to smartest and most careful. Every stage runs locally.
Nudge the transcriber toward the legal terms and case names in your active Pack before it even guesses.
Known fixes applied by exact rule. No model, no guessing, the same input always gives the same output.
A small on-device model resolves the hard cases, under the guards above, and every change comes back to you as a diff.
The app dictates fully offline with no account at all. An email gets you the Pack library, curated legal vocabulary kept current, plus updates and the LexpertAI ecosystem. Your email is the one Channel A item, and nothing else rides along with it.
One email. No transcript, no audio, no client data attached, that is the whole point.
Passwordless sign-in. Token lives in your macOS keychain. Offline after first login.