Is it allowed to clone someone else's voice?

Only with that person's consent and with rights to the recordings. SuperClonica requires both for orders and for published voices, and the Supertonic 3 license prohibits impersonating people without consent.

How long a recording do I need for voice cloning?

For training in the studio, one clean reference of 5 to 10 seconds with its exact text. Clean sound matters more than length.

Can I clone a voice without a GPU?

Yes, by ordering it from an author, who trains it on their own hardware. Training it yourself requires an NVIDIA GPU; using the finished voice does not.

How much does voice cloning cost?

Training it yourself is free, and the JSON downloads free of charge. An order costs the budget you set, from $0.01 to $10,000, and the author receives it only after you accept the result.

GUIDE

Voice cloning: how to get a voice that sounds like a recording

Updated October 23, 2026

Voice cloning means making a text-to-speech voice that sounds like the person in a particular recording. Some models do it on the fly from a few seconds of audio. On SuperClonica the voice is trained once into a Supertonic 3 file, which then speaks any text offline on an ordinary CPU.

Below: whose voice you may clone, how to record a good reference, the two ways to get the voice (order it from an author or train it yourself) and what the result can and cannot do.

Hear ready voices first Install the engine for your voice

Rights and consent come first

Clone your own voice, or the voice of someone who has agreed to it. When you order a voice, you confirm that you have the rights to the recordings and the speaker's consent to create a voice from them. Authors may publish voices only from recordings they have rights to and the speaker has agreed to.

The Supertonic 3 license, OpenRAIL-M, also lists impersonating people without their consent among prohibited uses. In practice: no voices of celebrities, colleagues or relatives without their permission, and keep that permission together with the source recordings.

How to record a good reference

The voice learns from what it hears, noise and room echo included. A short clean recording is worth more than a long noisy one.

Record in a quiet room with soft furnishings: no music, fans, traffic or echo.
One speaker only, with no overlapping voices or sound effects.
Keep the same distance to the microphone and a steady volume, without distortion on loud words.
Speak the way the voice should sound: a calm narrator reads calmly, a character performs.
Write down the exact text of the recording, word for word, including repeats and slips. The site asks for it with every reference.
Length: 5 to 10 seconds per reference for training in the studio. For an order you can upload WAV, MP3, FLAC or OGG files of up to 30 MB and 10 minutes each, up to 20 files.

Option 1: order a voice from an author

Without an NVIDIA GPU, or without the time to train, place an order in the voice commissions section.

Describe the task: timbre, manner of speech, what the voice is for and how you will judge the result. Set the voice language, a budget from $0.01 to $10,000 and a deadline in days.
Add the recordings with their exact text, confirm your rights and publish the order.
Authors respond with proposals, and you choose who does the work.
Reserve the budget on your balance. The author starts after that and delivers a trained voice with an audio demo.
Accept the result or ask for specific changes. The author is paid only after you accept, and the voice then downloads as a JSON file. If the order is cancelled, the reserved money returns to your balance.

Option 2: train the voice yourself

Training in the studio is free if you have a computer with an NVIDIA GPU. It becomes your worker: you connect it with a personal token, and it trains only your voices.

Create a voice and add a reference: a clean recording of 5 to 10 seconds with its exact text. References stay private.
Press Train on the reference, choose the number of steps and start.
Compare checkpoints by their demos and similarity, and continue training the most promising one.
Download the JSON free of charge on the training page. If you like, publish the voice for sale with a cover, a description and a price.

What to expect from a cloned voice

Training finds a timbre and manner of speech close to your recording. Expect a recognizable voice rather than a perfect copy: the cleaner the reference and the closer it is to the sound you want, the closer the result. A voice made to order is private, and the author does not publish it in the catalog.

The result is a file that runs offline at full speed. By our estimate SuperClonica speaks 6.4–9.7× faster than real time on a desktop 8-core CPU, and every line starts from the same timbre. The voice works with the 31 Supertonic 3 languages; Chinese is not available yet.

If you need a clone from a short clip right now, zero-shot models do that without training: Qwen3-TTS from 3 seconds of audio, OmniVoice from 3–10 seconds, Chatterbox from a reference clip. Their licenses and hardware needs differ, and our comparison pages cover both.