Private dictation
A plain-English confidentiality guide for professionals · Last updated: 1 August 2026
If you're a doctor, lawyer, therapist, accountant, or anyone else with professional secrecy obligations, you need to know exactly where your dictation goes before you speak a patient's or client's details near it. This page is that answer — plainly, completely, and without marketing gloss.
Where your dictation actually goes
That's the whole journey. There is no UtterKey server in the middle: we never receive, store, or see your audio or your text — ever.
What UtterKey (the company) can see
- Your purchase: name, email, and licence details from the payment processor
- Licence checks: your licence key and an anonymous device identifier, checked periodically so your licence works
- Nothing else. No analytics, no tracking, no dictation content, no audio — there's no infrastructure that could receive it even by accident
What the AI provider sees
The provider you choose (Groq or OpenAI) processes your audio to produce the transcript, and the transcript to produce the cleaned text — on your own account, under your own agreement with them. Both providers state that data submitted through their APIs is not used to train their models. Their retention windows and enterprise options change over time, so verify the current terms directly — that's part of your professional due diligence, and it takes five minutes:
- OpenAI: "How we use your data" (official API documentation) — at the time of writing: no training on API data; limited-time retention for abuse monitoring; zero-data-retention arrangements exist for eligible organisations — ask them
- Groq: privacy policy and trust & security portal (security attestations live there) — states API data is not used for training; check current retention terms
What stays on your machine
- Your API key and licence key — in the operating system's secure keychain, never in a plain file
- Your templates, snippets, and vocabulary — local files on your computer
- Your settings backup, if you export one — and it deliberately never contains your keys
Setting up for maximum confidentiality
1. You remain the data controller
Dictating patient or client information means you're processing personal data — same as typing it into the same system. UtterKey changes the input method, not your obligations under whichever rules govern you — HIPAA, GDPR, state or national privacy law, or your profession's own code. Your organisation's usual policies about cloud services apply to your choice of AI provider.
2. Choose the provider deliberately
Review your chosen provider's current API terms (not their consumer-app terms — API terms are typically stricter in your favour). Larger organisations can often sign data-processing agreements or zero-retention arrangements directly with the provider.
3. Minimise identifiers where practical
Long-standing dictation practice applies unchanged: many clinicians dictate "the patient" and let the letter template carry the demographics. UtterKey's templates make this easy — the identifying fields live in your document; your voice fills the clinical content.
4. Use your own hardware discipline
UtterKey activates per device (up to 3). Keep it on machines you control, with the OS-level protections your profession already requires — disk encryption, screen lock, managed accounts.
5. On shared machines, turn off "Show latest result" (Professional)
Need-to-know applies inside an organisation too: a ward or office computer may be used by several people, and none of them should read another's dictation. Untick Show latest result on UtterKey's Home screen and the app retains nothing to display — your text lands in your document and nowhere else. With the setting off there is nothing stored for anyone to reveal, even by switching it back on afterwards.
For organisations: don't trust us — verify
If you're evaluating UtterKey for a practice, chambers, clinic or firm, you shouldn't have to take a vendor's word for any of this. Our privacy claim splits into two parts — one your own IT people can verify with their own eyes, and one your lawyers cover with contracts they already know how to sign.
Part 1 — watch the wire yourself
UtterKey has no server in your audio path, and that is empirically checkable: run the app behind your firewall or proxy for an afternoon and watch its traffic. Below is the complete list of every address the app ever contacts. If you observe anything else, we want to hear about it.
| Address | What is sent |
|---|---|
| api.groq.com or api.openai.com | Your audio and transcript — to whichever provider you chose, on your API key. This is the only place dictated content ever goes. |
| utterkey-license.lambent.workers.dev | Licence key and an anonymous device identifier, for activation checks. Never any dictated content — the messages are a few hundred bytes; your proxy can confirm. |
| utterkey.com | Update checks (a version number) and template-library downloads. No content, no audio. |
Part 2 — cover the AI provider with paper, not hope
Nobody can "watch the wire" inside OpenAI or Groq — organisations cover that side the way they cover every cloud service: contractually. Because the API key is yours, you contract with the provider directly — UtterKey is not in that relationship. The instruments your counsel will recognise: OpenAI's API terms (no training on API data), its signable Data Processing Agreement, Zero Data Retention arrangements for eligible organisations, a BAA for US healthcare, and SOC 2 audit reports on request. Groq offers its own no-training commitment and security attestations — and if your risk team prefers one provider's paperwork over the other's, switching is a dropdown in the app, because the choice of provider is your control, not ours.
The sentence that reframes the assessment
For your dictated content, UtterKey-the-company is not a data processor at all — audio travels from your machine to your own AI account under your own agreement, and we never receive, store, or have any ability to access a word of it. The vendor-risk assessment you would normally run on a dictation supplier therefore collapses to almost nothing; the meaningful assessment targets the AI provider — an entity you can bind with real contracts. Compare the usual architecture, where every dictation transits the vendor's own cloud and your confidential material is one startup's security posture away from exposure. We removed ourselves from the blast radius by design.
Questions from your DPO, IT security, or compliance reviewer are welcome — hello@utterkey.com — including requests for this information in a form that slots into your vendor-assessment, HIPAA review, or DPIA paperwork — whichever framework your organisation runs.
What about fully-local AI?
Some organisations want dictation where the audio never leaves the building at all. We're actively exploring support for locally-run AI engines for exactly this audience. No promises and no dates on this page — if it matters to your organisation, email hello@utterkey.com and tell us; interest genuinely shapes the roadmap.
This guide is a factual description of how UtterKey works, not legal advice. For your specific obligations, consult your organisation's data-protection officer or professional body.
UtterKey for doctors → · for lawyers → · for professionals →