Text to Speech — Free, No Limit, WAV Download

Text to Speech

Paste any amount of text, pick a voice, listen — and when you like it, download a WAV. Three engines: your system's built-in voices with nothing to download, Kokoro 82M for Chinese and Supertonic 3 for English and Japanese. Everything is synthesised in the browser; your text stays on your device.

Your text stays in this tab. It is never uploaded, saved or sent with analytics.

Text

0 charactersLanguage

The text never leaves this page and is not saved when you close it.

Engine

Details
Uses the voices your operating system already has, so it sounds different on every computer. It only plays: the page never receives the audio, so there is nothing to download.
The voices belong to your operating system, and its terms apply.
Listen to this text with the built-in voices first; download the model when you want an audio file.
Details
License: Apache-2.0
Version: 6cc0f0d
Model weights download from Hugging Face, pinned to one version: onnx-community/Kokoro-82M-v1.1-zh-ONNX
The Chinese pronunciation dictionary downloads from this site (bytescope.dev): 9.3 MB
The Chinese voices speak Mandarin with a Beijing accent. Some characters with more than one reading come out wrong, and English words inside Chinese sentences can sound unnatural. Traditional characters are converted to Simplified internally to look up the reading; your text itself is not changed.
With WebGPU it is much faster than real time. Without a usable GPU it runs on the CPU at about 1.5–3× the length of the audio, so long text means waiting until it is synthesized.
While it runs, the browser uses about 2.5–3 GB of memory. Many phones cannot manage that.
Listen to this text with the built-in voices first; download the model when you want an audio file.
Details
License: OpenRAIL-M
Version: 3cadd1e
Model weights download from Hugging Face, pinned to one version: Supertone/supertonic-3
English and Japanese only. It does not speak Chinese.
With WebGPU it is much faster than real time. Without a usable GPU it runs on the CPU at about 1.5–3× the length of the audio, so long text means waiting until it is synthesized.
While it runs, the browser uses about 2.5–3 GB of memory. Many phones cannot manage that.
OpenRAIL-M use restrictions (Attachment A). You agree not to use the model or what it produces:
- (a)in any way that breaks an applicable national, federal, state, local or international law or regulation;
- (b)to exploit or harm minors in any way, or to try to;
- (c)to generate or spread verifiably false information or content meant to harm others;
- (d)to generate or spread personally identifiable information that can be used to harm someone;
- (e)to generate or spread content and place it in any context (for example a bot that posts) without stating clearly that it is machine generated;
- (f)to defame, disparage or otherwise harass others;
- (g)to impersonate, or try to impersonate, someone without their consent (for example deepfakes);
- (h)for fully automated decisions that adversely affect a person's legal rights or create or change a binding obligation;
- (i)to discriminate against or harm people or groups based on their social behaviour or known or predicted personal characteristics;
- (j)to exploit the vulnerabilities of a group (age, social, physical or mental traits) to distort someone's behaviour in a way that causes or is likely to cause physical or psychological harm;
- (k)to discriminate against people or groups based on legally protected characteristics;
- (l)to give medical advice or interpret medical results;
- (m)to generate or spread information used to administer justice, law enforcement, immigration or asylum processes, such as predicting that someone will commit fraud or a crime.
This is a summary, not legal advice. The license text linked above is what counts.

Voice and speed

VoiceSpeed 1.00×

Listen

To get an audio file, switch to a downloadable voice.

Sentences

Sentences appear here once there is text and a language.

Downloaded voices

checking…

What it can't do

Built-in voices can't be downloaded, and they sound different on every computer because they come from the operating system.
Kokoro's Chinese has a Beijing accent, misreads some characters with several readings, and English words inside Chinese may sound unnatural.
No voice cloning, and no emotion or tone control.
Supertonic 3 is under OpenRAIL-M, which has use restrictions; they are listed under its Details.
Without a GPU, synthesis takes about 1.5–3× the audio length, so you wait until it is synthesized.
The downloadable engines use about 2.5–3 GB of browser memory while they run.

About this tool

Most text to speech sites stop you at a few hundred characters, ask for an account, or send your text to a server to be read out. This page does none of that. The speech is generated inside the browser tab: there is no character limit and no upload — a long article, a whole chapter or a script you have not published yet never leaves your computer. Long texts are synthesized in parts of 10,000 characters. You listen in the page first, and the downloadable engines write a 16-bit mono WAV file when you ask for one. Once a voice model has been downloaded, the page keeps working without a network connection, which makes it an offline text to speech tool as much as a TTS in the browser.

Three engines, chosen by language

The built-in engine uses the Web Speech API (speechSynthesis), so the voices are whatever your operating system ships — they differ between Windows, macOS, Android and iOS — and there is nothing to download. Kokoro 82M handles Chinese with 100 voices; it is released under Apache-2.0 and needs a one-time download of about 340 MB for the model plus a 3–4 MB Chinese pronunciation dictionary. Supertonic 3 handles English and Japanese with 10 voices, downloads about 400 MB the first time, and is released under the OpenRAIL-M licence. Supertonic 3 does not speak Chinese, so for Chinese text the choice is between Kokoro and the built-in voice.

Listen first, download when you are happy

Every engine plays in the page before anything is written to disk. While it reads, the sentence being spoken is highlighted, and you can click any sentence to start from there; the speed slider runs from 0.5× to 2.0×. When the result sounds right, the Download WAV button renders the whole text to a file. The built-in engine has no download button, and the FAQ below explains why — the audio goes straight from the browser to your speakers and the page never receives it.

What to expect from the Chinese voices

Kokoro's Chinese voices have a Beijing accent, misread some characters that have more than one pronunciation, and give English words inside a Chinese sentence a noticeable accent. Traditional characters are converted to simplified internally only to look up pronunciations; the text you see is not changed. The built-in voice reads Chinese with whatever voice your operating system provides, so its accent depends on your system — at the cost of not being downloadable.

Speed and memory

With WebGPU available, synthesis is many times faster than real time. Without a usable GPU the models run on the CPU, and long texts then take longer than their playing time — in testing, about 1.5 to 3 times the audio length — so a ten-minute reading is a wait, not a click. While a downloaded model is synthesising, the browser uses roughly 2.5–3 GB of memory, which many phones cannot spare; on a phone the built-in voice is the practical choice.

What this page does not do

There is no voice cloning, no emotion or style control, and no MP3 export — the output is WAV. Model weights are fetched from pinned versions on Hugging Face only after you agree to the download, and the Chinese dictionary is served by this site. Opening the page triggers no download and asks for no permission. Audio quality is what these open models produce; judge it by listening to a sentence of your own text before downloading anything.

Frequently asked questions

Why can't I download audio from the built-in voice?

Because the Web Speech API only plays. speechSynthesis hands the text to the operating system, which sends the sound directly to your speakers; the web page never receives the audio data, so there is nothing for it to save. To get a WAV file, use Kokoro (Chinese) or Supertonic 3 (English, Japanese), which synthesise inside the page and can write the result to a file.

Why does the Chinese voice sound like Beijing Mandarin?

Kokoro 82M's 100 Chinese voices all speak Mandarin with a Beijing accent. It also misreads some characters that have more than one pronunciation, and English words inside a Chinese sentence come out with an accent. Traditional text is converted to simplified characters internally to look up pronunciations — what you typed is not altered. The built-in engine uses your operating system's Chinese voice instead, so its accent depends on your system, but that voice cannot be downloaded.

Is my text uploaded?

Not by this page. Kokoro and Supertonic 3 run entirely in the browser, so the text never leaves your device. The built-in engine passes the text to your operating system's speech service; on some systems that service may synthesise in the cloud, and the browser and the OS decide that, not this page — we cannot see or control it. For text you want to be certain stays local, use one of the downloaded engines.

Can I use the audio commercially?

That depends on the licence of the model that produced it, not on this page. Kokoro 82M is released under Apache-2.0. Supertonic 3 is released under OpenRAIL-M, which carries use restrictions — for example, not impersonating others, not using the output to harm people, and labelling published audio as machine-generated; the page lists the terms before the download. The built-in voices are governed by your operating system's terms. Read the relevant licence yourself before publishing or selling anything; this page does not give legal advice.

Does it work without an internet connection?

Yes, once a voice has been downloaded. Kokoro and Supertonic 3 are stored by the browser after the first download, and synthesis runs locally, so the page keeps working offline. The built-in engine depends on your operating system: some system voices are fully local, others may need a network connection, and the page has no way to tell which.

Why is the first download so large?

Because you are downloading the voice model itself, not a list of voices. Kokoro 82M is about 340 MB plus a 3–4 MB Chinese pronunciation dictionary; Supertonic 3 is about 400 MB. That is the price of synthesising in the browser instead of on a server: the model has to be on your machine. The download starts only after you agree, the weights come from pinned versions on Hugging Face, and the browser keeps them so you download once.

Why is a long text slow?

Speed depends on your hardware. With WebGPU, synthesis is much faster than real time. Without a usable GPU the model runs on the CPU, and in testing a long text took about 1.5 to 3 times its audio length to render, so a half-hour reading needs patience. Synthesis also uses roughly 2.5–3 GB of browser memory, which is why phones may fail or run out; the built-in voice has no such cost.