Talkie — privacy policy

Last updated: 21 September 2026.

The short version

Talkie records what you say, sends it to a transcription endpoint to be turned into text, and keeps the result on your device. Whoever wrote Talkie receives none of it, because there is nowhere for it to go: there is no Talkie server, no Talkie account, and no analytics of any kind in the app.

The rest of this page says exactly what that means, because "we don't collect your data" is a sentence every app says and almost none of them can spell out.

What Talkie is

Talkie is a client for a protocol, not a service. It talks to an endpoint that implements a widely implemented HTTP protocol for language and transcription models, and which endpoint that is, is a setting. The app ships with an address already in that field so that setting Talkie up asks you for one thing rather than three — but it is a setting, you can change it on the same screen where you enter your key, and everything the app does follows whatever the field holds. Point it at a different service, or at something you run yourself, and Talkie works exactly the same way.

The credential is always yours. You obtain it, you enter it, and you can remove it.

What stays on your device

All of it, unless it is on the list in the next section.

Your recordings. Audio is written to a file inside Talkie's own storage while you speak, and stays there afterwards.
Your transcripts, in every stage Talkie keeps — what came back from transcription, what the correction pass produced, and anything a template shaped. These live in a database on your device, and at the same time as plain Markdown files in a folder you can open without Talkie, because your writing should outlive the app.
Anything you gave Talkie to read aloud — a pasted paragraph, a fetched page, a shared article, a file you opened — and the highlights you made in it. These are kept in the same database and written out as Markdown in the same folder. One paragraph at a time is sent away while it is being read, and only in the case described in the next section.
A folder of your own files, if you pick one to read from. Talkie reads it where it is and never copies it anywhere. The only thing it writes there is == around a paragraph you asked it to highlight, in that file, on that press — the extended-Markdown highlight convention, so the mark is in your notes rather than only in Talkie. It creates, deletes, moves and rewrites nothing.
Everything you configure: templates, vocabulary, the folder Talkie saves into, retention settings, preferences.
Your statistics, which are computed on the device from your own captures and never leave it.

There is no sync, no cloud backup operated by us, and no account to attach any of it to. If you back your device up yourself, your backup contains what your device contains; that is between you and whoever operates your backup.

What leaves your device, and where it goes

There are two kinds of destination, and they are different enough to be listed apart: the endpoint in your settings, which is the only address Talkie has of its own, and an address you hand over yourself, which Talkie fetches and nothing more.

To the endpoint in your settings. Five requests, all over HTTPS, all on one transport, and this is the complete list.

A request for the endpoint's model list. No audio, no text of yours.
The transcription request — your audio file, the identifier of the model you are using, a language hint if you set one, and your vocabulary list, which is sent to help the model spell names it would otherwise get wrong.
A text request, for correction and for any template you apply — the transcript text, your vocabulary list, and the instructions for that stage.
A speech request, when something is read aloud in the endpoint's voice — one paragraph at a time of whatever is being read, the identifier of the speech model, and the voice name. This is the one request that can carry the text of your own documents: if you point Talkie's read-aloud at a folder of your notes, the paragraphs it reads are sent to the endpoint to be turned into audio. **You decide whether that happens, in Settings → Read aloud → Voice and speed, on the first row of the page: Read with.** It has two settings. Talkie's voice is the one Talkie ships with, and it is the one described above — the endpoint's speech model whenever your endpoint has a key and offers a model that reads text aloud, falling back to this device when it does not. This device uses the voice built into your device for everything, sends no part of any document anywhere, and is not reversed by anything that happens later: adding a key, regaining a network, or a speech model appearing in your endpoint's catalogue will not put your documents back on the network once you have asked for this device. The device's voice sounds worse, and that is the whole of the trade.

Two things worth knowing rather than discovering. Reading aloud has always worked with no key at all, synthesising on the device and sending nothing — so this setting is what makes that state reachable without giving up transcription, which is what it cost until 21 September 2026. And a document you have already heard is replayed from audio kept on your device, in whichever voice it was read in, so a paragraph you heard before you changed this setting may still sound like the endpoint. Nothing is sent for it; the audio is already there. Storage → Read-aloud audio is where that is cleared.

A route check when you save an endpoint or a key — an empty JSON object, sent to each candidate transcription path until one answers, so that Talkie can find the right one instead of asking you for it. No audio and no text of yours.

Your API key travels with each of these as an authorisation header. It goes nowhere else, is stored nowhere else, and is never written to a log or included in an error.

Talkie will not send your audio, your transcripts or your key over plain HTTP. The endpoint must be an HTTPS address; the check happens before a request object exists, so there is no path on which any of that could travel unencrypted.

To an address you hand over yourself. Talkie can read a web page aloud. When you paste a link, share one to Talkie, or press the read-aloud shortcut over a copied address, Talkie fetches that page — one GET, to the host in the address you gave it, following redirects to wherever that address leads.

It carries nothing of yours. No API key, no cookie, no stored credential, no identifier, and nothing from your transcripts or your notes. The session is a throwaway one with cookie and credential storage switched off, so nothing a page sets is kept and nothing is carried to the next one.
It happens only when you hand over an address, on that press, once. Talkie does not watch your clipboard, crawl, prefetch, or follow links inside a page it has fetched.
The address itself is what leaves. The host you name learns that somebody asked it for that page, as any web request does.
This one is not held to HTTPS, because the address is yours: if you hand Talkie an http:// link it will fetch it as given. Nothing of yours is in that request, but the address is in the clear.

And that is the whole list. There is no host Talkie contacts on its own initiative, no fallback host, no analytics endpoint, and no diagnostic upload. Every address above is either the one in your endpoint setting or one you typed, pasted or shared in the moment before the request was made.

What the endpoint does with it is between you and them

This is the part worth reading twice. When your audio and your text reach the endpoint, they are under whatever agreement you have with whoever operates it — their privacy policy, their retention period, their jurisdiction, their decisions about training. We are not a party to that arrangement, cannot see what happens, and cannot make promises on their behalf.

If that matters to you, and for some recordings it should, read the terms of the service in your endpoint setting, or change the setting to a service whose terms you prefer, or to something you run yourself. The app does not care which you choose.

Talkie does name two services, and it is worth being exact about what that is and is not. The screen where you enter your key offers two ready-made endpoints by name, with a link to the page that explains where each one's key comes from, and a Custom… choice beside them for your own address. They are there because the alternative — describing the shape of a service you should go and find — is accurate and useless to somebody setting the app up for the first time.

Naming them is a recommendation, not an endorsement. Talkie is not affiliated with either, receives nothing from either, and is not paid if you choose one. Nothing in the app works only for them: every route is derived from whatever the endpoint field holds, and an endpoint speaking the same protocol works whether or not it is in that list. We do not compare services, rank them, or claim a relationship we do not have — and the two names appear on the screens that help you get started and nowhere else. Nothing in the store listing names a provider.

What we collect

Nothing.

No analytics, no telemetry, no crash reporting, no attribution, no advertising identifier, no device identifier.
No accounts, no sign-in, no email address, no profile.
No third-party SDK that could collect anything on our behalf. Talkie has one external dependency, a database library, which makes no network calls at all.
No contact with any server we operate, because we operate none.

This is not a policy choice that could quietly change in a background update. There is no server to send anything to, so there is nothing to switch on.

Your API key

Held in your device's keychain. It is marked as available only after the device has been unlocked once, and explicitly marked not to synchronise, so it does not travel to your other devices or into an encrypted keychain backup. It is never stored in Talkie's database, never written to a file, and never included in an error message shown to you or recorded anywhere.

It is sent as an authorisation header to the endpoint in your settings, and to nowhere else.

You can remove it in Settings at any time, in one action. Removing it stops Talkie recording, which is the intended behaviour rather than a failure.

Permissions Talkie asks for

The microphone, because the app records what you say. This is the only permission the mobile app requests.
On macOS only, Accessibility, and only if you want Talkie to paste a finished transcript into whatever you are typing in. It is optional, Talkie explains what it is for before the system asks, and declining leaves the transcript on your clipboard instead. You can decline permanently.
On macOS only, folder access, when you save transcripts to a folder you picked, or when Talkie creates its own Markdown folder inside Documents.

Talkie asks for no camera, no photo library, no contacts, no location, no calendar, no local network, and no speech recognition permission — transcription happens at the endpoint, not on the device.

The one thing stored outside your own data

A single true-or-false value, recording that a one-time onboarding screen has been shown, so it is not shown again. It is stored in the app's own preferences, contains nothing about you, and is removed when the app is.

Children

Talkie is not directed at children and collects nothing from anyone, so it collects nothing from children either. It has no social features, no user-to-user content, and no advertising.

Deleting things

A single transcript: delete it in the app. Its Markdown file goes with it.
Everything: remove the app. Its container, database, audio and settings go with it. The keychain item is removed with the app's data.
Before you delete anything, you can export your entire history as files in one action, because data you cannot take with you is data you do not really have.

There is nothing to request from us, because there is nothing of yours on our side to request.

Changes to this policy

If this policy changes, the date at the top changes with it. A change that made any of the above less true would be a change to what the app does, not just to this page — and the whole point of the design is that we cannot make that change quietly.

Contact

Privacy questions: Open an issue on the support tracker