Medical dictation without the cloud

Every dictation vendor selling to clinicians answers the privacy question the same way: with a policy. There is a stronger answer available, and almost nobody offers it.

Ask a medical dictation company what happens to your audio and you will get a well-drafted paragraph. It is encrypted in transit. It is encrypted at rest. It is deleted after thirty days, or ninety, or on request. They are SOC 2 Type II certified. They will sign a Business Associate Agreement.

All of that can be true and all of it is still a promise about what someone else does with your patient's information after they receive it. Every one of those sentences describes a control on data that has already left your building.

What “HIPAA-ready” actually means

The phrase does a lot of quiet work in this market, so it is worth being precise about it, including where Walkie sits.

HIPAA does not certify software. There is no government body that inspects a dictation app and stamps it compliant. Compliance is a property of a covered entity's whole practice — its policies, its training, its contracts — not a badge a vendor earns. Any tool advertising itself as “HIPAA certified” is describing something that does not exist.

A BAA moves liability; it does not move risk. When a vendor signs a Business Associate Agreement they take on contractual obligations around the protected health information you send them. That is genuinely valuable — it is how the law expects cloud services to be used. But the PHI still goes to them. A BAA is the paperwork that makes the transfer permissible, not a mechanism that makes the transfer private.

“HIPAA-ready” means the technical controls are in place. For Walkie, that means cloud requests run under Zero Data Retention — the audio is transcribed, the text comes back, and the content is discarded rather than logged — and that nothing is ever used to train AI models, on any plan, contractually, with the providers in the chain. It does not mean certified, and Walkie does not sign BAAs.

The third answer

Walkie runs dictation two ways, and the difference is the entire point of this article.

Cloud Mode

Audio goes to a transcription provider, comes back as clean, punctuated, filler-free text formatted for the app you are dictating into. It is faster and it writes better. It runs under Zero Data Retention and nothing trains on it. It is a well-protected transfer — and it is still a transfer.

On-Device Mode

The speech model runs on your own machine. It makes no network request at all. Not a smaller one, not an anonymised one — none. The audio is captured, transcribed, and typed into your EHR without a packet leaving the computer. It works with the wifi off.

This is the mode where the compliance conversation ends early, because there is no business associate to have an agreement with. You cannot breach a database that was never written to. You cannot subpoena logs that were never generated. The failure modes that a BAA exists to assign liability for are not mitigated — they are absent.

What on-device costs you

Being honest about this is the only way the rest of the article stays credible. On-device transcription is worse than cloud transcription in specific, predictable ways:

• Punctuation is basic. You will get sentences, but the cloud model is noticeably better at commas, paragraphs, and knowing when a list is a list.
• Filler words stay in. The cloud pass strips the “um”s and false starts. On-device does not.
• No app-aware formatting. Cloud Mode adapts the output to where it is going. On-device gives you the text.
• It uses your machine. A larger, more accurate model costs you memory and a beat of latency. You pick the tradeoff from several models.

Your custom dictionary works in both modes — corrections are applied locally after transcription — so drug names, colleagues' names, and the abbreviations your specialty actually uses come out right either way. In practice that closes most of the accuracy gap that matters clinically.

Dictate the note; think hard about recording the visit

These are different features with genuinely different privacy properties, and conflating them is how people get surprised.

Dictation in On-Device Mode is the airtight case described above. Nothing is stored beyond your local history, which you can clear.

Meeting recording is stored, by design. A searchable archive of past conversations is, definitionally, a stored archive. Walkie's recordings and transcripts live in your account until you delete them — and deletion is permanent, covering the rows, the audio, and the cloud copy. On a Mac with Apple Intelligence, Walkie can capture and summarise a meeting entirely on-device, which keeps the content local. Recorded hours are metered the same either way, so choosing on-device capture is a privacy decision and never a cheaper one.

A short decision rule

Use On-Device Mode when the words themselves are the protected thing: clinical notes, therapy session write-ups, anything naming a patient and a diagnosis in the same sentence. Use Cloud Mode for the writing where polish matters more than secrecy — internal drafts, admin email, the third rewrite of a policy document.

You switch between them whenever you like. It is a per-moment privacy choice, not a plan you are locked into, and both modes exist on the free tier and on every paid tier.

Questions