Voice Cloning Online

What is AI voice cloning?

Voice cloning online, as implemented here, takes a short recording of someone speaking and uses it to generate new speech that sounds like that same voice, saying words they never actually said. This tool does it with zero-shot cloning, meaning it needs only one short reference clip rather than hours of recorded training data — you provide 5 to 15 seconds of clean speech, and the model uses that single sample to shape the voice of anything you generate afterward.

Under the hood, this runs on Chatterbox, an on-device voice cloning model that currently supports English. Because it runs locally in your browser through WebGPU, with a WebAssembly fallback on devices without WebGPU support, your reference recording is never uploaded anywhere — the model downloads once to your device and does all the work there.

Getting a good reference recording

The quality of your reference clip has more influence on the result than almost anything else in this process. Aim for 5 to 15 seconds of a single person speaking clearly, without background noise, music, or other voices layered in. The tool checks your recording automatically and flags it if it is too quiet, clipped (peaking and distorting), or outside the expected length, so you get quick feedback before generating anything.

• Record in a quiet room, close to the microphone, without background noise or music playing.
• Speak naturally at a normal pace and volume — avoid whispering, shouting, or exaggerated delivery.
• Avoid clipping: if the recording sounds distorted or the level meter peaks into the red, record again at a lower input level.
• Stick to a single speaker in the clip. A reference with overlapping voices or background conversation confuses the model.
• A little variety in the reference — a full sentence or two rather than a single flat word — tends to give the model more to work with.

The model download and device requirements

The first time you use voice cloning, your browser downloads the Chatterbox model. It is a large download — over 1 GB — so the first use takes noticeably longer than later ones, particularly on a slower connection. After that first download, the model is cached locally, and subsequent clones on the same browser and device start without downloading it again.

Because this is a demanding on-device AI model, a desktop browser with WebGPU support — Chrome or Edge are recommended — gives the smoothest experience. It will run without WebGPU using a WebAssembly fallback, but expect it to be considerably slower. Mobile devices can struggle with both the download size and the processing demands, so a desktop machine is the more reliable choice for this particular tool.

Consent and ethics: only clone voices you have permission to use

Voice cloning technology makes it possible to generate speech in someone else's voice, and that capability carries real responsibility. Only clone a voice you own — your own voice — or a voice you have explicit permission to clone from its owner. The permission checkbox in the workflow exists to make that expectation unavoidable before you generate anything.

Do not use voice cloning to impersonate someone without their consent, to create audio intended to deceive or defraud, or for any use where the person being cloned has not agreed to it. This applies whether the source voice is a public figure, a colleague, a family member, or anyone else — public availability of someone's voice online is not the same thing as permission to clone it.

What to expect: limitations of voice cloning

Zero-shot cloning from a short clip is genuinely capable, but it has real limits worth knowing before you rely on it. Accent and emotional expressiveness are the two areas where fidelity varies the most — a reference recorded in a flat, neutral tone will generally produce flatter-sounding generated speech than one with more natural inflection, and a distinctive accent may not always carry through with full accuracy. Longer or more expressive scripts can also drift slightly from the reference character over the course of the generation.

This tool currently supports English. If your reference recording or your intended script is in another language, results are likely to be unreliable or unavailable. As with any generative audio tool, treat the output as a strong approximation of the reference voice rather than an indistinguishable copy — small artifacts and inconsistencies are normal, especially on longer scripts.

Privacy: your reference recording stays on your device

Everything in this workflow — the reference recording, the model itself, and the generated speech — stays on your device. Nothing is uploaded to a server at any point, whether you record your reference through the microphone or upload an existing file. The only network activity involved is the one-time download of the Chatterbox model itself, which is a public model file, not a transfer of your voice or your script anywhere.

Voice cloning vs. text to speech: what is the difference?

It helps to know when to reach for which tool. Text to speech generates natural-sounding narration using one of more than 20 built-in voices — fast, no reference recording needed, and well suited to narration, accessibility, or any script where the specific voice does not need to match a real person. Voice cloning exists for the opposite case: when you specifically want the output to sound like one particular voice, built from a short reference recording of that voice, with all the consent responsibilities that entails.