A text-to-speech app that never uploads your documents

Most reading apps describe themselves as private and mean a promise: a policy page saying your text is not retained, or is deleted after processing, or is not used for training. Those can all be true and they are all still promises about what a company does with something it received.

Sandbook's version is narrower and checkable. The voice is generated on your iPhone, so the text being read never has to go anywhere. Opening a book, extracting its text, recognizing a scanned page, generating speech and exporting audio all happen on the device. It is not a retention policy. It is the absence of the upload a retention policy exists to govern.

There are two ways to check that without taking our word:

• Turn on airplane mode and read a book. Import an EPUB or a PDF, press play, and listen to the whole chapter with the radios off. Speech that is generated on a server cannot survive that test; speech generated on the phone does not notice it. The only thing that waits for a connection is adding a new web page.
• Scan a page with the radios off. Point the camera at a printed page with airplane mode on. The text is recognized and read aloud anyway, because recognition runs on the device too.

The formal version lives in the privacy policy. This page is the engineering account of why your documents stay put.

Everything you read, is the short answer. The longer one names the parts, because “on device” is a phrase that has been worn smooth by marketing.

The model and the voices. Kokoro-82M, open-weight neural TTS with 82 million parameters, is bundled inside the app download along with the 41 voices; the download is about 276 MB with all of it included. Nothing unpacks from a server on first launch, there is no activation step and there is no API key. That is why the download is large, and it is the whole reason the rest of this page is possible.

Your files. A book you import, whether EPUB, PDF, TXT, RTF, Markdown or HTML, is copied into Sandbook's own sandbox directory. Deleting it from the library removes that copy and does not touch your original.

Your scans. Pages you scan with the camera, and photos or screenshots you pick, are recognized on the device with Apple's own text recognition. The recognized text becomes a book in your library, one chapter per page. The camera is used only while you scan.

Your text. Extracted chapters, PDF text layers, typed text, pasted text and clipboard text are processed in memory and spoken from there. The audio is synthesized on the GPU through Apple MLX at 24 kHz, sentence by sentence as you listen. The same model also runs in a browser tab on a computer, if you want to test the voices on a private sentence first; that text stays in the tab as well.

Your reading position and voices. Where you stopped in each book, the voice and speed each book remembers, your starred voices and your reading options sit in local app preferences. There is no iCloud sync and no cross-device handoff, which is a genuine limitation, and also the reason none of it is anywhere else.

Your exports. M4A audio you export goes wherever you send it through the iOS share sheet, and nowhere else.

Web pages you add. When you add a page from the library's + menu or share it from Safari, Sandbook fetches that page, and its cover image or icon, directly from the site that hosts it, like a browser would. No Sandbook server receives the address or the article. Once it is in your library it reads offline like any other book.

A page about privacy that lists only benefits is an advert. The bill for this architecture is real and it is paid in features.

• Real-time reading pauses when the screen locks. The model holds real GPU memory while it runs, and iOS suspends that work when the app leaves the foreground. To keep listening with the screen locked, prepare a chapter's audio ahead of time first.
• Six voice languages, not dozens. 41 voices, 28 of them US and UK English, plus Spanish, French, Italian, Brazilian Portuguese and Hindi. A hosted reader can offer far more.
• iOS 18.2 and an iPhone 12 or later, or an iPad with an A14 chip or later. A hosted reader runs on older hardware, precisely because it is not running a neural model on your phone.
• No text recognition for image-only PDF files and no DOCX import. Camera scans and photos are recognized; a scanned PDF file is not.
• A large download, and no iCloud sync, no account recovery and no library that survives the device.

If any of those matter more than the upload question, say so early and pick differently: the alternatives page compares the readers honestly, including other on-device readers that do several things Sandbook cannot. What running a model on the phone means in general is on the local TTS page.

One generic fact separates the two architectures, and it does not require anyone to behave badly for it to be true: if the speech is synthesized on a server, then the text reached that server. What happens to it afterwards is a policy question that every vendor answers for itself. Whether it was transmitted at all is not a policy question; it follows from where the model runs.

So the useful thing to check on any reading app is not the privacy page. It is the architecture. Each of these is the vendor's own description of where its speech is generated, read on 14 September 2026:

• Speechify. Mostly on Speechify's servers. On 18 December 2025 Speechify added on-device iOS voices that run locally through Apple's Core ML.
• NaturalReader. The AI voices are generated on NaturalReader's servers; the Free Voices tier uses the device's own pre-installed system voices.
• ElevenReader. On ElevenLabs' servers. Audio is downloaded for offline use rather than synthesized on the device.

Note what that list does and does not say. It makes no claim about any competitor's document handling or retention; for that, read each vendor's own privacy policy. The point is narrower: where the voice is made decides whether your text has to travel, and each vendor states that in its own words before you install anything.