Every voice in Sandbook

Every voice here is a style embedding for the same model, Kokoro-82M, running on the device's GPU at 24 kHz. There is no basic tier and no premium tier, no “AI voices” charged separately from “standard voices”, and no voice that needs a connection to speak. The 41 voice packs ship inside the app, which is part of why the download is large.

That is the honest form of the “most natural sounding” claim. Among text-to-speech apps that run entirely on the iPhone and cost nothing, these are the most natural voices we know of, not the best voices on any phone, and not a match for a hosted service with a thousand of them and a monthly bill. Press play on a few below and settle it before you install anything. Want to hear your own sentence instead of ours? Try the voices online, in this browser, free.

What differs between these voices is language and character, not quality: how warm, how fast, how formal, how much the reading leans into the sentence. A voice that suits a thriller is rarely the one you want for a specification document, so the app is built around trying them rather than reading a list.

Press play on any voice below. These are the same Kokoro-82M model and the same voice packs the app ships, speaking the line Sandbook uses for its own in-app auditions, in each voice's language: “Take care of people and the world, and make good new things.”

Nothing is fetched until you ask for it: the page ships no audio at all until you press a play button, and each clip is about four seconds and 30 KB.

The most natural sounding English voices for fiction are the expressive ones: Heart, Bella, Kore and Isabella carry dialogue and shift tone across a sentence. Fable and Onyx suit anything with weight to it: history, horror, a long-form essay you want read seriously.

Non-fiction and documentation are easier to follow in something even and unhurried, where you notice the argument rather than the performance: Adam, Alloy, River, Emma and Eric hold a flat, clear line for an hour without wearing on you. If you listen at higher speeds, the calmer voices stay intelligible further up the range than the animated ones do.

For anything you're studying, try the same paragraph in two voices before committing to a 400-page book. It takes thirty seconds and saves several hours.

The English catalogue is 20 American voices and 8 British ones. Sandbook 1.0.4 added 13 more: Dora, Alex and Santa in Spanish; Siwis in French; Sara and Nicola in Italian; Dora, Alex and Santa in Brazilian Portuguese; and Alpha, Beta, Omega and Psi in Hindi. French has one voice today, so if French is your main language, you are choosing Siwis.

The voice decides the language. To listen to a Spanish book, give that book a Spanish voice; the book remembers its voice and speed, so the next English book you open keeps its own. The clean-up that reads dates, prices, units and acronyms naturally applies to English text only.

The app itself is available in English, Spanish, French, Italian, Portuguese (Brazil) and Hindi. There is no voice cloning, no custom voice creation and no pronunciation dictionary, and there are no voices outside these six languages. The 41 voices on this page are the complete set, and they are the same 41 for everyone.

Speed runs from 0.5× to 2.5×, and it is applied inside the model by scaling the predicted duration of each phoneme rather than by speeding up a finished waveform. That is why a voice at 2× still sounds like that person reading quickly, with its pitch and character intact, instead of the chipmunk effect you get from resampling audio.

Each book remembers its own voice and speed. The samples above are all at 1×.