Text to speech on an iPhone XS, and other older iPhones

An iPhone XS was announced in September 2018. It is eight years old, it still gets iOS updates, and the question people arrive with is whether a modern neural voice — the kind that sounds like a person rather than a satnav — will run on it at all, or whether “on-device AI” is a phrase that quietly means “buy a new phone”.

When this was written, it ran. But three separate gates decide it, only two of them are printed on the App Store listing, and different apps set them in very different places.

Note what the last column does and does not say. On 14 September 2026, every app here bottomed out at the same 2018 hardware at most, because Apple’s own iOS 17 and 18 device lists both stop at the XS and the XR — the difference between these apps is not which phones they can reach but which ones they have chosen to. Éist’s iOS 17 floor is the friendliest to somebody who has stopped updating; Local Narrator’s iPhone 15 line rules out five years of phones that could run the model.

“On-device AI” has come to mean a several-billion- parameter language model, and those really do need a recent phone and several gigabytes to sit in. Kokoro-82M is not that. It has 82 million parameters, and Sandbook's whole download is about 276 MB with the model and all 41 voices included, which is two orders of magnitude smaller than the local chat models a current iPhone is being sold on.

What it needs from the hardware is a GPU generation modern enough for the Metal feature set Apple MLX targets, and enough spare memory to hold the weights while the rest of the app runs. At the time of writing, the 2018 chip cleared both. (With 1.0.4, Sandbook found that its MLX build needs the GPU family introduced with the A14, and raised its floor to iPhone 12.) It is genuinely the case that the constraint here is the software stack’s baseline rather than the arithmetic — which is why an eight-year-old phone can produce a voice that would have needed a data centre when it was new.

There is one thing an older phone does not get, and it is not audio quality: it is time. Generation speed scales with the GPU, so the same chapter is synthesised more slowly on an older chip than on a new one. Because Sandbook generates sentence by sentence with a look-ahead buffer and starts playing after the first sentence rather than after the chapter, what that mostly costs you is a slightly longer pause at the very start. I have not published per-device timings and I am not going to guess at them here; the architecture is the part I can describe honestly, and the App Store refund window is the part that settles it for your specific phone.

On a supported device, from iPhone 12 up, there is still no reduced tier: every device gets the same 82 million parameters, the same 24 kHz output, all 41 voices (28 of them US and UK English), the same 0.5× to 2.5× speed range applied inside the model, word-by-word highlighting and the same per-book position.

And if your phone is below that floor, an iPhone XS, XR or 11, or an iPhone X, 8 or 7 that stopped at iOS 16, the answer is the one already on the device. Speak Screen is part of iOS Accessibility, costs nothing, and works on far older releases than any app here. Speak Screen against a neural reader covers what you trade for that.

The product page for this, which iPhones run Sandbook, has the full iPhone and iPad list in one place.