How to use Text to speech
- Type or paste English text, up to 3,000 characters.
- Pick one of 13 American or British English voices and a speed from 0.8× to 1.5×.
- Select Read aloud. The first time, the voice model (about 56 MB) downloads once; then playback starts as soon as the first part is ready.
- Listen again with the player, or download the speech as a WAV or MP3 file.
A natural voice without sending your text anywhere
Hearing a draft read back catches clumsy sentences, and a recorded voice is handy for videos, slides, lessons and accessibility. This tool reads English text aloud in a natural-sounding voice, plays it as it is made, and lets you keep the result as an audio file.
Online voice generators send your text to their servers. This one runs a speech model inside your browser instead, so a script that isn't public yet, internal training notes or a private letter stay on your device, and WeaveAI never sees them.
How it works
The voice comes from Kokoro, an open speech model with 82 million parameters released under the Apache 2.0 licence. We use an 8-bit version of it, with a few layers left at 16-bit, run by ONNX Runtime Web in a background worker in this tab, on the processor.
Before the model can speak, each word becomes a string of speech sounds. That step uses misaki, the pronunciation library written for Kokoro: its American and British dictionaries of about 184,000 and 197,000 words, its rules for plurals, past tenses and -ing forms, and a small spelling model for names and words the dictionary lacks. Numbers, money, percentages, times and years are spelled out first, so $1,250.50 or 10:30 come out the way people say them.
Long text is read a few sentences at a time. Playback starts as soon as the first part is ready and the rest follow in order. On a recent laptop, each part takes roughly two to three times as long to make as it lasts, so short pauses can appear between parts; the finished file has none.
Voices, formats and limits
There are 13 voices: nine American English (six female, three male) and four British English (two female, two male), each with its own accent's dictionary. Speed goes from 0.8× to 1.5×. Kokoro also has voices for other languages, but they depend on a pronunciation engine under a licence we can't ship, so this tool is English only.
The result plays in the page and downloads as a WAV file (24 kHz, 16-bit mono) or as a 128 kbps MP3, made with the same audio engine as the video tools. Texts can be up to 3,000 characters, about three minutes of speech. Unusual names may be mispronounced, so listen before you publish.
Questions
- Is my text uploaded?
- No. The voice model is downloaded to your browser, and the text is turned into speech on your device. The only download is the model itself, from WeaveAI's own engine server; your text never leaves your device.
- How big is the download, and how long does it take?
- The model, one voice and the English dictionary come to about 56 MB, downloaded once and then kept by your browser. Making the speech takes roughly two to three times as long as the audio lasts on a recent laptop, and longer on phones, so playback may pause between parts while the next one is made.
- Which languages and voices are there?
- English only: 9 American English voices and 4 British English voices, each reading with its own accent's dictionary. Kokoro has voices for other languages, but they need a pronunciation engine whose licence we can't ship, so they are left out.
- Can I use the audio commercially?
- The Kokoro model and its voices are released under the Apache 2.0 licence, which allows commercial use, and the audio is made on your own device. You are responsible for the text you have read out.