Free AI Voice Cloning — In Your Browser, Nothing Uploaded

Record about eight seconds of someone speaking, type whatever you want, and hear it said in their voice. The unusual part is where it happens: the model runs inside this page on your own hardware, so the recording of the voice never leaves your device. Almost every other voice cloning service works by uploading a sample of a person’s voice to a server — which is a meaningful thing to hand over, and the reason this one was built the way it was.

How to Free AI Voice Cloning — In Your Browser, Nothing Uploaded

  1. 1

    Download the model

    The first visit fetches about 1.5 GB and caches it in your browser. It only happens once, and afterwards the tool works with no connection at all. This is by far the largest download here, and it is the honest price of running a real voice cloning model locally rather than sending your audio to someone else’s server. Three of the model’s four parts are only published uncompressed, so there is no smaller version to offer.

  2. 2

    Record or upload a voice

    About eight seconds of clear speech from a single speaker, with no music or background noise behind it. There is a sentence on the page to read aloud that covers a good spread of sounds. Anything past twenty seconds is trimmed, because the model takes a fixed-size impression of a voice and extra audio only costs time.

  3. 3

    Type what it should say

    Up to 2,000 characters. Punctuation does real work — full stops and commas are how the model places pauses and shapes intonation, so properly punctuated text sounds noticeably better. Longer text is split at sentence boundaries and joined back together automatically.

  4. 4

    Adjust expressiveness and generate

    The expressiveness slider controls how much the delivery performs, from flat and even up to dramatic. Past about 1.0 it starts drifting away from sounding like the person you recorded, so it is worth a listen at the default first. Then play it back and download it as an MP3 or a lossless WAV.

How it works

The voice comes from Chatterbox, a zero-shot text-to-speech model released by Resemble AI under the MIT licence — including the weights, which matters more than it sounds: several of the best-known open voice cloning models publish permissive code alongside checkpoints licensed for non-commercial use only, so the audio you make with them is not really yours. Zero-shot means there is no training step. A speech encoder listens to your reference clip once and compresses that voice into a fixed set of numbers; a language model then predicts speech tokens for your text conditioned on those numbers, and a decoder turns those tokens into a 24 kHz waveform. Because the impression of the voice is fixed-size, a two-minute reference is no better than an eight-second one. All four parts run through WebGPU, your graphics hardware, which is what makes a model this size practical in a browser tab at all — it generates a little faster than the speech it produces.

Supported formats

WAV, MP3, M4A, AAC, OGG and FLAC. Files are decoded by your browser, so anything it can play will work here. Output is exported as WAV.

Frequently asked questions

Is my voice recording uploaded anywhere?
No. The model is downloaded to your browser the first time you use the tool, and from then on the recording goes straight into it on your own device. You can prove it: load the page once, turn off your Wi-Fi, and cloning still works. This is the whole reason the tool exists in this form — a voice sample is a biometric, and uploading one to a service is not something to do casually.
How much audio do I need to clone a voice?
Around eight seconds of clear speech is the sweet spot, and three is the practical minimum. More than about twenty seconds does not help: the encoder compresses whatever it hears into a fixed-size impression, so a long clip produces the same amount of information about the voice while taking longer to process. Quality of the clip matters far more than quantity — one speaker, no music, no room echo, no background chatter.
Why does it need WebGPU, when the other tools do not?
Because it is far bigger than they are, and the older WebAssembly path cannot carry it at a speed anyone would tolerate. Pointed at that path in testing, the model was still loading after four and a half minutes without having produced any audio, on a machine where WebGPU has it ready in under two and generating faster than realtime. Rather than ship a fallback nobody could sit through, the page checks for WebGPU and says plainly when it is missing instead of starting a 1.5 GB download that would disappoint you. Recent Chrome, Edge and Safari all have it; Firefox support depends on your platform.
Can I use the cloned audio commercially?
The model permits it — Chatterbox is MIT licensed, weights included, with no revenue ceiling and no separate registration. Whether you may clone a particular voice is a different question, and it is about the person, not the licence: using someone’s voice generally needs their permission, and several jurisdictions now specifically protect voice likeness. Your own voice, or one you have clear permission to use, is the safe ground.
Is the audio watermarked?
No. The reference implementation of this model passes its output through an inaudible watermarker that marks the audio as machine-generated, but that step is not part of the browser build, so audio generated here carries no such mark. Worth knowing in both directions: it means nothing is embedded in your audio, and it also means you should not rely on this tool to produce clearly-labelled synthetic speech where a disclosure obligation applies to you.
What languages does it work in?
English. Chatterbox has a multilingual version covering twenty-three languages, but it cannot yet be run through the browser library this tool uses, so what is here is the English model. It will clone a voice with an accent perfectly well — it is the language of the text being read that is the limit.
Does it use Chatterbox Turbo?
No, and deliberately. Turbo is newer, less than half the size and about twice as fast, so it was the obvious choice — but its published browser build is broken in a way that is easy to miss: it returns roughly 2.7 seconds of speech and then pads the rest of the audio with silence, no matter how much text you give it. Every input tested produced an identical 2.7 seconds. The standard model has no such limit, generating a clean half-minute in one pass, so this tool uses that instead and pays for it in download size.
How long does it take?
A little quicker than the speech it produces, so around six seconds of waiting for a nine-second line — measured on a mid-range laptop, and slower on an older one. The slow part is the one-off 1.5 GB download. After that, capturing a voice takes about half a second and only happens once per voice, not once per generation.
Why does the cloned voice not sound exactly like me?
Zero-shot cloning takes an impression of a voice rather than learning it, so it captures timbre, pitch and general character well and finer habits of speech less so. The most common fixable causes are the reference clip: background noise, room echo, more than one person talking, or too little speech in it. A clean eight seconds recorded close to the microphone makes a bigger difference than anything else on the page.
Does this use AI?
Yes. It runs Chatterbox Turbo, a neural text-to-speech model from Resemble AI, which conditions its output on a short recording of a voice rather than on a fixed set of built-in ones. The difference from a hosted voice cloning service is only where the model runs, which here is your own machine.
Do I need an account?
No. The free tools work straight away with no sign-up. Creating a free account raises your daily processing limit from 20 to 60 minutes.
Is it free?
Yes. Every tool on this page is free to use, with a daily limit on total audio processed. There is no watermark and no trial period.

Other free audio tools