Cloning a voice: what sample you need, and whose voice you may clone

September 15, 2026

Cloning a voice has become trivial. That's exactly what makes it a problem, because the technical question — can I? — stopped being the interesting one a while ago. The interesting one is the other.

Let's take both.

The sample: less than you think, cleaner than you think

For a clone to come out well, the reference file has to meet four conditions:

Between 6 and 12 seconds. No more. Below six you lose tonal depth; above twelve it doesn't improve proportionally and only adds latency when loading it. The intuition that "the more audio the better" is false here.

A complete sentence, with natural intonation. Not a single word, not a monotone reading. The model is capturing how that voice rises and falls, and it needs to see it happen.

Zero background noise. This is what most people get wrong. No music, no effects, no room echo, no air conditioning. Any foreign frequency contaminates the output voice and will show up in everything you generate afterwards. Ten seconds recorded in a closet full of clothes gives a better result than a minute recorded in a living room.

16-bit or 32-bit float WAV, at 22,050 Hz or above. 44,100 is better. Mono or stereo doesn't matter, it's normalized automatically.

The detailed guide, with the complete flow from any source to the final sample, is in CLONE_VOICES.md inside uttera-tts-hotcold.

And now the real question

A voice is an attribute of an identifiable person. In Spain it is expressly protected — Organic Law 1/1982 names it alongside image and name — and in the rest of Europe the reasoning is equivalent.

That means there is a short list of things you may do and a long list of things you may not.

You may, with peace of mind:

• Your own voice. To dub yourself into a language you don't speak, so you don't have to re-record every time a sentence changes, to have a version of yourself available at three in the morning. This is the good case and it's the one most people need.
• The voice of someone who gave you written permission, for a specific use. A hired voice actor with an express assignment; a colleague who agrees to voice the internal assistant.
• A voice that doesn't exist, generated or licensed for the purpose.

You may not, even though it technically works:

• The voice of an actor, a presenter or a politician because public material exists. A video being accessible is not a license over the voice of whoever appears in it.
• The voice of a recognizable person to make it look as though they said something they didn't. At that point we're no longer talking about likeness rights, we're talking about impersonation.
• The voice of a customer or an employee because you have it recorded from a call. Recording for quality assurance is not collecting a voice model.

The acid test

If you're in doubt, one question usually settles it: would the person that voice belongs to be comfortable seeing what you're about to generate with it?

It isn't a legal standard, but it's right almost every time. And when the answer is "well, they'll never find out", you already have your answer.

How we handle it

The reference voices for cloning are not included in the published code, and that's deliberate: it isn't material that can be handed around. You supply your own samples and register them; no code to touch.

And when you clone a voice on the fly through the API, that audio never enters the cache, whether you ask for it or not. In the most sensitive case, the default decision shouldn't depend on the customer remembering to ask. The post about the voice cache covers it in detail.

Anything to add or correct? Write to support@uttera.ai. If you correct us, we edit the post and credit you.

← All posts