What VV can and cannot protect

Last updated 23 September 2026

The threat model, in plain English

Security pages tend to say "we take security seriously" and stop. This one lists who could see what.

Someone with your laptop

If a person has your unlocked computer, they can see whatever VV has kept: the history of what you have dictated, and any audio VV is still holding, until you delete it. Two things limit that. First, the retention setting: you choose in Settings how long audio is kept, from three days to never deleting them, and you can clear history at any time. Second, the usual system protections apply, because VV stores nothing outside your user account; FileVault on a Mac, or BitLocker or Device Encryption on Windows where it is turned on, and your login password protect VV's data the same way they protect your documents.

Someone watching your network

No audio and no dictated text, because recognition happens on your computer and there is no network path for either. They would see the app download the speech model on first run; register an anonymous identity with Google Firebase at first launch, and again whenever you sign in; report the number of words in a dictation, so the free allowance can be counted; send anonymous product analytics, which are on unless you turn them off in Settings; check the licence; and, only if you have set one up, send the text of a cloud transform to the provider you chose, or, in Pro, the text you send to Cloud Polish or for meeting summaries. All of those connections use TLS.

Vaitly, the company

If you have an account, we know the email address on it and which plan it holds. From every install, including free ones with no account, we receive anonymous usage counts tied to a device identity: roughly, how many words were dictated this week, so the free allowance of 1,000 words can be enforced. We never receive audio, or the names of the apps you dictate into. On Free, there is no mechanism in the app that sends dictated text unless you set up a cloud transform yourself. Pro changes one thing: when you use Cloud Polish or meeting summaries, that text passes through our server to OpenAI. Audio never does.

Where things are stored

History, audio and the speech model live in VV's own data folder in your user account: under Library/Application Support on a Mac, under AppData on Windows. They are ordinary files on your own disk, so they inherit whatever protection the disk has: FileVault on every Apple Silicon Mac, BitLocker or Device Encryption on a PC where it is on, and your login password. Nothing is stored in iCloud or OneDrive, nothing is synced between devices, and removing the app and that folder removes everything.

Audio retention is yours to set

Settings → Advanced (Auto-Delete Recordings) lets you choose how long audio recordings are kept: three days, two weeks, three months, or never deleted. Out of the box it keeps only the latest five recordings, so you can replay a recent dictation and correct the text without a pile building up. There is no setting that keeps none at all: the most recent recording exists until the next one replaces it, or until you delete it. History entries can be deleted one at a time or cleared in bulk. This is the single most useful control for the "someone with my laptop" case, so if that worries you, set retention to three days and clear the history.

The one optional boundary: cloud transforms

Transforms rewrite text after it is recognised: tidy this up, make it an email, translate it. By default they stay on your computer: Apple Intelligence on a Mac, or a local Ollama model on a Mac or PC. You can instead configure a cloud provider with your own key. If you do, three things hold: only text is sent, never audio; the transform is labelled as cloud in the app so you know before you use it; and nothing is sent unless you invoke that transform. Standard dictation is unaffected. If you never configure a provider, no text leaves your computer for any reason.

Pro features and their boundary

Pro adds four things that involve Vaitly's server, on Mac and Windows. Vaitly Meetings is in beta. Each is narrower than it could have been, and none of them sends audio.

• Vaitly Meetings recording. Calls are detected on your computer from which app is using the microphone. No bot joins, and nothing is installed in or granted to Teams, Zoom or any other service, so no admin or tenant permission is involved. Microphone and call audio are transcribed on your computer and never written to disk; transcripts are stored locally, and each recognised line is journalled locally so a crash loses nothing.
• Meeting summaries and questions about your meetings. (The weekly meetings pack is assembled on your computer from summaries you already have and sends nothing.) The transcript text goes to Vaitly's API, never to OpenAI directly, and the app holds no provider key. The server checks your account and plan, sends the text to OpenAI with storage off, and logs counts, never content. Replies are checked against the transcript: an owner nobody named is removed, and an item carrying a number the meeting never mentioned is dropped. Zero data retention is not in place with OpenAI.
• Cloud Polish. The same path for the finished text of a dictation, with a monthly spending ceiling and a check on the reply before it is returned.
• Encrypted backup. A 256-bit key is created on your computer and, on a Mac, kept in the Keychain. Each item is compressed and encrypted with XChaCha20-Poly1305 before upload, bound to your account and its storage name, and sent over TLS 1.3 — so it is encrypted at rest and in transit, and storage names are keyed hashes that reveal nothing. Vaitly also holds a copy of that key, wrapped so that it is released only to a machine signed into your account: that is what makes a second machine pick your snippets and transcripts up with nothing for you to do. Being able to release it means Vaitly could in principle open it, and we would rather say so than claim otherwise. If you prefer the stronger guarantee, turn on a recovery key: Vaitly’s copy is then wrapped with a key only you hold, nobody but you can open the backup, and a new machine will ask you for that key before your data appears on it.

Code signing and notarisation

Every Mac build of VV is signed with an Apple Developer ID certificate issued to Vaitly Limited and submitted to Apple for notarisation, with the hardened runtime enabled. When you open the app, macOS verifies the signature and the notarisation ticket before it runs. This means the app you launch is the one we built, has not been altered on the way to you, and cannot load unsigned code into itself. If Gatekeeper warns you about a copy of VV, that copy is not ours: download again from the download page and tell us.

The Windows installer is Authenticode-signed by Vaitly Limited. Windows checks that signature and shows the publisher name before it installs; it installs for your user account only, with no administrator prompt. If Windows names any other publisher, or none, that copy is not ours.

Updates

Honest answer: there is no in-app updater. Updates are downloaded from the download page and installed the same way as the first copy: dragged to Applications on a Mac, or by running the new installer on Windows. Settings, vocabulary and history are kept. The version you are running is shown in the app; the current releases are 1.0.0 for Mac and 1.1.0 for Windows.

What the licence check sends

The app holds a signed licence and an anonymous device identity. When it checks in, it sends those and receives back whether the plan is valid and how many words remain on the free tier. That is the whole exchange: no dictated content, no audio, no file or app names. A paid plan covers 5 devices, and the device identity is what lets you move it between them from Settings → Account.

Reporting a security problem

Email letstalk@vaitlyvoice.com with what you found and how to reproduce it. We acknowledge every report within one working day, keep you informed while we work on it, and thank you in the changelog if you would like. Please give us a reasonable window to fix it before you publish. We do not run a paid bounty programme at this stage.

What we are not

VV has no SOC 2, ISO 27001 or HIPAA certification, and we do not claim any. We have not yet published a penetration-test report. We make no claim that VV is suitable for medical, legal or other regulated work; whether it meets your obligations is a judgement for you and your organisation, and we would rather you made it with the facts above than with a badge. As those things change, this page will change with them.

Limits and the privacy boundary

Nothing on this page makes VV "completely secure", because nothing is. What it does is keep the sensitive part, your voice and the text it becomes, on the computer you control, and make every exception visible and optional. The privacy policy covers the legal side; on-device recognition covers the technical one.

Compared with cloud dictation

A cloud dictation service has to protect your audio on its servers, in transit, and in the hands of its staff and subprocessors, and you have to trust that it does. VV's approach is to have no audio to protect on any server in the first place. See how VV compares with the alternatives.