A voice of her own
Clone a voice from a single reference clip and bind it to a character for good. No training run, no dataset, no waiting.
Watch a clip become her voice
The whole trade is on the right: a short reference goes in, and a voice that belongs to one character comes out.
- Bound to her for good. The voice rides the character record — live sessions, TTS calls and rendered clips all speak through it.
- The transcript is handled. Whisper reads the reference for you, so you never type out what was said in it.
- The voice she uses on camera. In a live session, her replies are synthesised through this clone.
About ten seconds of me talking, recorded once, is all it takes.
And now I can say the rest myself — for good.
From one clip to a permanent voice
Upload one reference clip
A single clean sample of the voice you want. No studio session, no scripted read, no hours of recording — the engine is zero-shot, so it works from a short sample rather than a training set.
The clip is transcribed for you
Whisper transcribes the reference automatically, so you do not have to type out what was said in it. That transcript is what the cloning engine aligns against.
The voice is registered to a character
The reference is registered into the synthesis container and bound to a character record. From then on, anything that character says routes through that voice.
She speaks in it, everywhere
The same voice serves live sessions, text-to-speech calls and rendered clips. There is no separate "cloned voice" product to wire up.
What you get, precisely
Zero-shot, not fine-tuned
No training run, no queue, no waiting for a model to converge. The reference conditions synthesis directly.
GPT-SoVITS · port 8509Preset voices for everything else
Four neural presets in US and UK English cover narration, product and character work when a clone is not needed.
Piper · port 8506Word-level timings
Synthesis returns per-word timings alongside the audio, so lip movement and subtitles run off the same clock. Alignment is interpolated, not forced — treat the timings as close, not frame-exact.
Approximate alignmentDegrades instead of failing
If the cloning container is unavailable, synthesis falls back to a preset voice rather than returning an error mid-conversation.
Automatic fallbackStreaming synthesis
Long text can be synthesised in chunks and streamed, so playback starts before the whole passage is rendered.
Chunked endpointSelf-hosted end to end
The reference clip and every generated sample stay on your infrastructure. No third-party voice API receives either.
Local containers
What is shipped and what is not
Cloning works through the dashboard and the API. Two things the engine does not do are listed too, so you do not design around them.
| Path | Status | Detail |
|---|---|---|
| Cloning in the dashboard | Shipped | Upload, auto-transcribe, register and synthesise all work end to end. |
| Preset text-to-speech API | Shipped | Generate, stream and request word timings over the public API. |
| Creating a clone over the API | Shipped | Submit a reference URL and the job runs: fetched, transcribed and registered. Poll the voice until it reports active, or failed with a reason you can act on. |
| Emotion and style control | Not built | There is no emotion parameter. Delivery follows the reference clip. |
| Cross-language cloning | Not built | Presets cover US and UK English. A clone speaks the language of its reference. |
Whose voice you are allowed to clone
Consent is enforced by the terms you accept, not yet by a technical control. There is no automated consent verification or audio watermarking in the platform today. Do not describe either as a safeguard to your own users.
You may clone your own voice, or a voice you hold documented, revocable permission to reproduce. Cloning a person’s voice without their agreement — public figures included — is prohibited under the acceptable use policy, and accounts found doing it are terminated.
If you operate a platform where performers submit their own voices, you are the controller of that consent relationship. Keep your own records; the platform does not currently keep them for you.
Asked before the first clone
How much audio do I actually need?
One clean clip of roughly ten seconds. The engine is zero-shot — the reference conditions synthesis directly, so there is no dataset to assemble and no training run to wait for.
Can I control emotion or delivery?
No. There is no emotion parameter — delivery follows the reference clip. If you need a different mood, record a reference in that mood.
Does it work in other languages?
The presets cover US and UK English, and a clone speaks the language of its reference. Cross-language cloning is not built.
What happens if the cloning engine goes down?
Synthesis falls back to a preset voice rather than failing mid-conversation. Your integration keeps speaking; it just does not sound like her until the container is back.
Can I create a clone over the API?
Yes. Submit a reference URL and the job runs — fetched, transcribed and registered. Poll the voice until it reports active, or failed with a reason you can act on. Request shapes are in the API reference.
Whose voice am I allowed to clone?
Your own, or one you hold documented, revocable permission to reproduce. There is no automated consent check — the obligation is contractual, under the acceptable use policy, and accounts that break it are terminated.