Sandbook

Kokoro TTS online, free and private in your browser

Kokoro TTS online, free: type text or import an EPUB, PDF, Word or text file, pick one of 41 Kokoro voices in six languages, and hear it read aloud right here. Your text and audio never leave your browser.

In short

Can I use Kokoro TTS online for free?

Yes. This page runs Kokoro-82M, the open-weight text-to-speech model, inside your browser tab, free and with no account. It offers all 41 voices Sandbook ships, in English, Spanish, French, Italian, Brazilian Portuguese and Hindi. Typed text goes up to 1,500 characters at a time, with a WAV download; an imported EPUB, PDF, Word or text file of any length opens in a full-screen reading view and is read aloud from any sentence. Your text and audio never leave your browser, because the speech is generated on your own device.

Last verified November 11, 2026

How it works

  1. Type or paste your text, or import a file: EPUB, PDF, Word (.docx), Markdown, HTML or plain text.
  2. Pick a language and a voice. The play button next to the voice picker plays a recorded sample without loading anything.
  3. Press Start. Your browser downloads the Kokoro-82M model once and keeps it in its cache; from then on everything happens in this tab. The first sentence starts playing as soon as it is ready, while the next ones render behind it, and the result is a WAV to download.

An imported file is opened in this browser tab; nothing is uploaded. It opens in a full-screen reading view, with the chapters of an EPUB under Contents. Tap any sentence and the tool reads aloud from there; the sentence being read is highlighted, and you can pause or tap another one at any time.

Under the hood: sentences, phonemizers, the background worker

The model runs on your own computer. Your text is split into sentences, each sentence is turned into phonemes, and the model turns the phonemes into 24 kHz audio in the voice you picked. The inference runs in a background worker, so the page stays responsive.

English text goes through Kokoro's own English phonemizer, which also spells out numbers, prices and times. Spanish, French, Italian, Portuguese and Hindi go through eSpeak NG, the same phonemizer Kokoro uses for those languages, loaded only when you pick one of those voices.

For an imported file the tool generates only a sentence or two ahead of what you hear, so a whole book never has to be rendered at once.

The same engine is set up for one job each on four other pages: read text aloud, an online EPUB reader, read a PDF aloud and hear your essay read back.

Text to voice, text to audio: what this tool makes

If you came looking for a text to voice or text to audio tool, this is one. Type or paste up to 1,500 characters, pick one of the 41 voices and press Generate: the text becomes speech in your browser, and Download WAV keeps it as a 24 kHz mono file that any video or audio editor opens. For a script longer than one run, generate it in parts and join the files in your editor.

It is a free AI voice generator in the plain sense of the words: the voices are synthetic, there is no sign-up and no credit meter, and the Apache-2.0 license of Kokoro-82M allows commercial use. It does not clone a voice or imitate a person.

What to expect: WebGPU or CPU

The tool picks the engine for you:

  • WebGPU (graphics card): the full-precision model, a one-time 326 MB download. On a recent laptop, a 40-word paragraph takes about 3 seconds and the first sentence plays after about 1 second. Current Chrome, Edge and Safari on a computer have WebGPU.
  • CPU (WebAssembly): the 8-bit model, a one-time 92 MB download, used when WebGPU is not available. It is slower than real time: expect the first sentence after 10 to 15 seconds and 20 to 40 seconds for the same paragraph.
  • Phones: most phones download a 163 MB CPU model once; phones with little memory get the 92 MB one. It works on recent phones and is slower than on a computer, with pauses between sentences.

Either way, the second visit skips the download: the model loads from the browser cache in a few seconds.

Engine details: model precision, phone memory, interrupted downloads
  • There is no quality setting to get wrong: on WebGPU the tool always uses the full model, because the smaller quantized ones produce garbled speech there.
  • On the CPU, speed depends on your processor. On a computer, plan on about 1 GB of memory in the tab while it runs.
  • Phones run the CPU model in a lighter setup that asks the browser for less memory and speaks in shorter pieces. Speech is generated more slowly than it plays. Browsers whose tab closed while the engine was running also get the 92 MB model.
  • If the download is interrupted, it resumes where it stopped. Closing the tab right after the first download can occasionally stop the browser from finishing its cache write, in which case the next visit downloads the model again.

Six languages, 41 voices

The voice decides the language. The tool offers the same 41 voices as the Sandbook app, with the app's names and descriptions:

  • English: 28 voices, 20 American and 8 British, including Heart, Bella, Michael, Emma and George.
  • Spanish: 3 voices (Dora, Alex, Santa).
  • French: 1 voice (Siwis).
  • Italian: 2 voices (Sara, Nicola).
  • Brazilian Portuguese: 3 voices (Dora, Alex, Santa — distinct from the Spanish voices of the same names).
  • Hindi: 4 voices (Alpha, Beta, Omega, Psi).

The app's beta voices are here too, in French, Spanish, Italian, Brazilian Portuguese, Hindi and British English. Each is listed after its language's own voices, marked “(beta)”, and the tool never picks one for you. The beta set can change between updates.

To audition a voice without loading anything, press the play button next to the voice picker, or listen to all the Kokoro voice samples. In this browser demo, speed runs from 0.5× to 2.0× (the app goes to 2.5×), and a fast voice still sounds like a person talking quickly.

Download, usage rights and limits

  • Download. Every generation of typed text can be saved as a 24 kHz mono WAV file. MP3 and M4A export are not available yet. Imported documents are read aloud in the page and have no audio download.
  • Usage rights. Kokoro-82M is published under the Apache-2.0 license, which allows commercial use. The audio you generate is yours; Sandbook claims nothing over it.
  • Limits. Typed text: up to 1,500 characters per generation, about a minute and a half of speech. Imported files: any length; scanned PDFs (no text layer) and DRM-protected books cannot be read.
  • Phones. The tool works best on a computer. It also works on recent phones, more slowly, and an older phone may not have the memory for it; the tool says so before downloading when that looks likely. Reading stops when the phone locks or the tab goes to the background.
  • Privacy. Your text and audio never leave your browser. A file you import is opened on your device and is not uploaded. The only files the tool fetches are the model, the voice you pick, for non-English voices the phonemizer, and the reader for the file type you import.
Open-source credits and licenses

The tool is built on open-source software: Kokoro-82M (Apache-2.0), ONNX Runtime Web (MIT), Transformers.js (Apache-2.0), kokoro-js (Apache-2.0), eSpeak NG (GPL-3.0), PDF.js (Apache-2.0) and fflate (MIT). Siwis and the French beta voices are built on the SIWIS French speech corpus (CC BY 4.0).

Want this offline on iPhone or Mac?

This page is a demo of the model. Sandbook is the app: it runs the same Kokoro-82M model on iPhone, iPad and Mac, fully offline, with all 41 voices and no length limit. It reads EPUB books and PDFs (scanned ones too, which this page cannot read), web pages and (on iPhone and iPad) camera scans aloud, highlights each word as it is spoken, remembers where you stopped, and exports a chapter or a whole book as one M4A file. Free, with no account and nothing to buy. It needs an iPhone XS, iPhone XR or later on iOS 18.2, or a Mac with Apple silicon (M1 or later) on macOS 15.2.

iPhone, iPad and Mac · Free, no in-app purchase

Questions about Kokoro TTS online

Is this Kokoro TTS online tool really free?

Yes. There is no account, no sign-up and no limit on generations. Each generation of typed text takes up to 1,500 characters, and you can run as many as you like; imported files have no length limit. It costs nothing to run because the model runs on your own device, not on a server.

Can it read an EPUB, PDF or Word file aloud?

Yes. Press Import a file, or drop the file on the tool, and choose an EPUB, PDF, Word (.docx), Markdown, HTML or plain-text file. It is opened in your browser, never uploaded, and shown in a full-screen reading view: tap a sentence and the reading starts there, so even a whole book works. Scanned PDFs without a text layer and DRM-protected books cannot be read, and there is no audio download for whole documents.

Is this a text to voice generator with a file I can download?

Yes. Typed or pasted text, up to 1,500 characters per run, is turned into speech with one of 41 voices, and each result downloads as a 24 kHz WAV audio file. There is no sign-up and no limit on the number of files. WAV is the only download format, and the tool does not clone voices.

Can I use the audio I generate commercially?

Kokoro-82M is released under the Apache-2.0 license, which allows commercial use. Sandbook claims no rights over the audio you make here: download the WAV and use it in a video, a course or a product. The voices are synthetic, so no real person's voice is involved.

Why is it slow on my computer?

Without WebGPU the tool runs a 92 MB model on the CPU through WebAssembly, which is slower than real time: roughly 20 to 40 seconds for a 40-word paragraph on a fast laptop, with the first sentence playing after 10 to 15 seconds. A browser with WebGPU, such as current Chrome or Edge on a computer, uses the full model on the graphics card and does the same paragraph in about 3 seconds.

Does it work on iPhone or Android?

Yes, on recent phones, but more slowly than on a computer. Most phones download a 163 MB CPU model once (the browser keeps it for later visits) and run it in a lighter setup; phones with little memory, and browsers whose tab closed while the voice engine was running, get the smaller 92 MB model. Either way, speech is generated more slowly than it plays, so expect a wait before the first sentence and pauses between sentences. Keep the screen on and the tab in front while it works. An older phone may not have the memory for it: when that looks likely, the tool says so before it downloads anything and lets you try anyway or hear the recorded samples. On iPhone and iPad, and on Macs with Apple silicon, the free Sandbook app runs the same model natively, faster and offline.

Does it work offline?

Partly. After the first load, the model files are cached in your browser, and speech is generated in the tab with no server involved. Opening the page itself still needs a connection. For fully offline Kokoro TTS on an iPhone, iPad or Mac with Apple silicon, use the Sandbook app, which bundles the model.

Sources

  1. Kokoro-82M model card (hexgrad), with its voice list — Hugging Face · checked November 11, 2026
  2. Apache License 2.0, the licence Kokoro-82M is published under — apache.org · checked November 11, 2026