tl;dr
Start with the rule your organization already has, then pick tools that fit it. Local recognition keeps audio and text off servers, which removes one category of risk and leaves others: a lost laptop, a cloud backup, an open-plan office, and disclosure obligations you cannot delegate to software. Build two lanes so the sensitive one is a habit rather than a decision made under time pressure. Auri Voice can run both steps on your Mac, a local engine plus Auri Local cleanup, and that is a technical property, not a compliance certification.
Start with the rule, then pick the tool
Sensitive work already has rules: privilege, patient confidentiality, employment law, an approved-tools list. The tool decision follows from those, and a tool cannot change your obligations.
- Which categories of content are you handling: privileged, clinical, personnel, financial?
- Does your organization maintain an approved-software list?
- Are there contractual requirements, such as an agreement a vendor must sign?
- Where are drafts allowed to live, and for how long?
- Who signs off on a new tool, and what do they need from you?
Answer those first and the shortlist gets short fast. Many workplaces allow a local dictation app more readily than a cloud service, because the review question changes from data processing agreements to endpoint software.
Write down what you decided and who agreed. A one-paragraph note in your own files is what protects you if the question comes up a year later.
What local recognition solves, and what it does not
On-device recognition means audio is transcribed on your machine rather than uploaded. That is a real reduction in exposure and a narrow one.
| Risk | Local recognition helps? | What handles it |
|---|---|---|
| Audio sent to a server | Yes | On-device engine, cloud off |
| Vendor retention of transcripts | Yes, if storage is local too | Local history, delete path |
| Lost or stolen laptop | No | Disk encryption, screen lock |
| Transcript in a cloud backup | No | Backup scope review |
| Being overheard | No | Closed room, or typing |
| Disclosure and retention duties | No | Policy and legal advice |
Keep the two questions apart: where processing happens and where text is stored. An app can transcribe locally and still write searchable history into a folder that syncs.
Also check the fallback path: some apps quietly use a server when a local model is missing or unavailable for a language. Test with the network off to see what happens when the local path fails.
Honest limits
No dictation vendor, us included, can hand you compliance. Software provides technical properties. Compliance is a program made of policies, contracts, training, access control, retention schedules, and evidence.
- On-device processing is a control, not a certificate.
- Where an agreement with a vendor is required, that is a contract question, not a settings question.
- Your professional judgment about what to dictate cannot be delegated to a tool.
- Marketing language about privacy is not a substitute for your security review.
Auri Voice transcribes on your Mac by default and can keep cleanup there too with Auri Local. It is not a HIPAA certification, and it does not decide whether your workflow is acceptable in your jurisdiction or your firm.
Bring this to whoever owns the question: your privacy officer, general counsel, security team, or professional body, and get the answer in writing.
Build two lanes so you stop deciding under pressure
The failure mode for sensitive work is a judgment call made while you are already talking and behind schedule. Two prepared lanes remove the decision.
Sensitive lane (in Auri Voice, four settings and a room)
- Engine: a local one, checked in Settings > Voice engine. The default Fast Multilingual Dictation already is local; the check is confirming nobody switched it to Auri Cloud.
- Cleanup: Auri Local, set in AI settings (Apple Silicon), so the transcript never leaves for polish.
- History: Settings > Dictation > History. Turn off Save transcriptions for the engagement, or leave it on and use Delete all history when the project ends.
- Meetings: set Saved audio to Delete after notes are ready, under the gear in Meetings > Capture and storage.
- Room: door closed, headphones off (they make you speak louder), nobody in earshot. Draft only into approved apps and storage.
Normal lane
- Everything else: cloud polish, phone keyboard, whatever is fastest.
Make the wrong lane cost an extra step: a note taped near the camera, a calendar block labeled for the sensitive work, history cleared before you start.
Then rehearse the sensitive lane once when nothing is urgent. Confirm recognition works with the network off, insertion works in the approved app, and history behaves the way you expect.
Dictate without saying the identifiers
You can often get the drafting speed without putting identifiers into audio or text. Speak the structure and add the details by hand.
- Use initials, a case number, or a role instead of a full name.
- Leave a placeholder for dates of birth, account numbers, and identifiers.
- Dictate the reasoning and paste the facts from the source of record.
- Skip diagnoses and allegations aloud when a shorthand works.
- Fill placeholders at the keyboard in the approved system.
This habit also reduces the cost of a mistake. A draft that says the complainant and the March incident is far less damaging in the wrong folder than one with names and dates throughout.
It helps accuracy too. Names and long numbers are exactly what dictation gets wrong most often, so typing them removes both a privacy risk and an editing chore.
Device and disk hygiene
These checks take fifteen minutes and cover the failure modes local processing does not touch.
- Turn on full disk encryption and confirm the recovery key is stored properly.
- Set a short screen lock and require the password immediately.
- Review what your backups include, including any folder holding dictation history.
- Check clipboard managers, screen recorders, and note apps that sync by default.
- Decide a retention rule for drafts, then follow it.
- Be careful while screen sharing, since inserted text appears live.
Clipboard managers and screen sharing deserve extra care: a utility that keeps everything you copy will archive the paragraph you were careful about, and text dictated while sharing a window appears to the room as you speak.
Questions to bring to your reviewer
Security reviews go faster when you arrive with the questions already framed. These map to what a reviewer needs to sign off.
- Where does recognition run, and what happens with no network?
- Where is history stored, is it encrypted at rest, and how is it deleted?
- Which features call a server, and can they be disabled centrally?
- Which third parties are involved, and under what retention terms?
- What minimum OS version and permissions does the app need?
- What does the vendor state in writing, and how is it dated?
For Auri Voice, the short answers: Auri Voice transcribes on your Mac by default. The cleanup pass sends the transcribed text (never audio) to the provider you pick — Auri Cloud unless you change it; choose Auri Local in AI settings to keep both steps on the Mac (Apple Silicon). Dictation keeps working with the network off once an engine is downloaded. History stays on the Mac and can be turned off or deleted in Settings > Dictation > History. Apple Watch audio reaches the Mac encrypted, is transcribed there by a local engine, and is deleted after 15 minutes. The app needs macOS 15 or later.
Get the rest in writing from us and from any other vendor you are comparing. Then let the people who own the compliance program make the call.
FAQ
Does local dictation make me HIPAA or GDPR compliant?
No. On-device processing reduces where data travels, and compliance depends on your policies, contracts, training, access controls, retention rules, and documentation. Treat a local app as one control among many, and get sign-off from the people responsible for that program.
Can I dictate privileged material at all?
Many people do, with rules: local recognition, cloud features off, a private room, no client names spoken aloud where possible, and drafts stored only in approved systems. Confirm with whoever owns your confidentiality policy before you build the habit.
Is a phone keyboard safe for sensitive text?
Treat it as the wrong lane by default. Auri Keyboard defaults to cloud dictation on iPhone, with an On-Device option (English-only). If sensitive content has to happen on a phone, switch that first and skip the server-side features.
What about being overheard?
It is the risk local processing does nothing about, and often the most likely one. Open-plan offices, shared homes, cafes, and trains all carry your voice further than you think. Use a closed room, or type.
What should I bring to a security review?
Where recognition runs, what happens with no network, where history is stored and for how long, how deletion works, whether backups capture it, which third parties are involved, and what changes when a cloud feature is enabled. Written answers, not screenshots of marketing.