Local TTS, running on your iPhone

TTS is text to speech: software that turns written words into a spoken voice. Local TTS means the model that does the turning runs on your own device. Cloud TTS means it runs on someone else's computer: your text is sent over the network, a server synthesizes the audio, and the audio is sent back.

Both produce a voice, so from the outside they can look the same. The difference is where the work happens, and three things follow from that. With cloud TTS the text has to travel, a connection is required at the moment of speech, and every sentence costs the provider server time, which is why cloud readers count characters, minutes or hours. With local TTS the text stays where it is, the connection is irrelevant, and there is no running bill to recover.

A few years ago local meant the flat, robotic system voices. That changed when neural speech models became small enough to fit on a phone and fast enough to run there in real time. Local TTS is now a trade of breadth for independence, not of quality for privacy.

A phone is where reading happens in the least predictable places, which is exactly where a server dependency hurts most.

• Privacy by architecture. Your books, PDFs and scanned pages are not uploaded, because nothing needs them anywhere else. Everything stays on your device, and there is no account and no tracking. The long version is on the private text to speech page.
• Airplane mode, tunnels, bad signal. Speech made on the device does not notice the network is gone. A chapter opened on a flight sounds the same as one opened at home. More on that under offline text to speech.
• No per-character meter. When the voice costs nothing to produce, there is no reason to ration it: no daily quota, no monthly hours, no premium voice tier.
• No queue. There is no upload, no rendering job and no waiting for a file to come back. Press play and the first sentence starts as soon as it is generated.

The engine is Kokoro-82M, an open-weight neural text-to-speech model with 82 million parameters. The download is about 276 MB, with the model and all 41 voices included, so nothing is fetched on first launch and there is no API key or activation step.

Sandbook runs the model on the device's GPU through Apple MLX, Apple's framework for machine learning on its own chips. Text is split into sentences and each one is synthesized as you listen, at 24 kHz, so playback starts after the first sentence rather than after the whole chapter. Speed runs from 0.5× to 2.5×, and each word is highlighted as it is spoken.

The same Kokoro model runs locally in a desktop browser, too: try Kokoro TTS in a browser tab, where the speech is generated on your computer and nothing is uploaded.

There are 41 voices in six languages: 28 in English (20 American, 8 British), 3 in Spanish, 1 in French, 2 in Italian, 3 in Brazilian Portuguese and 4 in Hindi. They read EPUB, PDF, TXT, RTF, Markdown and HTML files, web pages, camera scans and photos (recognized on the device), and typed or pasted text. Text in the Reader tab can be exported as an M4A file.

Running the model on the phone is a choice with a bill, and it is paid in the following:

• A large download. The download is about 276 MB, with the model and all 41 voices included. Cloud apps are small because their voices live elsewhere.
• A newer device. Sandbook needs iOS 18.2 on an iPhone 12 or later, or an iPad with an A14 or M1 chip or later. If your iPhone is older, see text to speech on older iPhones.
• Real-time reading pauses when the screen locks. The model uses the GPU, and iOS suspends that work when the app leaves the foreground. To listen with the screen locked, prepare a chapter's audio first, then lock the phone and play it.
• Fewer languages than cloud services. Six voice languages, not dozens. A server can hold a catalogue no phone has room for.

Any app can say it runs on the device. The test takes a minute and does not require trusting anyone:

1. Open a book, a PDF or some pasted text in the app.
2. Turn on airplane mode, with Wi-Fi off as well.
3. Press play and listen to a full chapter, then try a voice you have not used yet.

Speech generated on a server cannot survive that test. Speech generated on the phone does not notice it. In Sandbook the only thing that waits for a connection is adding a new web page, since the page has to be fetched from the site that hosts it. Scanning a printed page with the camera works in airplane mode too, because the text recognition also runs on the device.