AI Voice Cloning

Voice cloning makes new speech in the sound of one particular voice: its pitch, its texture, its accent and the way it shapes words. Earlier systems typically needed long studio recordings and a training run. Many current models work from a short reference clip: they listen to a few seconds of the voice and speak new text in a voice that matches it.

On Mage, a cloned voice starts as an audio reference. You upload an MP3 or WAV, pick a creation or a saved sound, or paste an audio or social link, choose where the clip starts, and Mage keeps up to 15 seconds under an @handle. Mention that handle in a Seed Audio prompt, and it speaks your text in that voice as an MP3 of exactly the length you set. There is no training step: the clip is read each time you use it.

The same voice carries into video. Save a clip as a character’s voice and Cherry, Raspberry, Plum, Lemon and MiniMax-H3 use it whenever that character speaks; Cherry, Plum, Lemon, MiniMax-H3 and Blueberry also take audio references in a video prompt. Clone only your own voice, or a voice you have clear permission to use.

The model copies what it hears, including the room. A clip recorded in a kitchen with an echo comes back sounding like a kitchen. Record in a small room with soft furnishings, a hand’s width or two from the microphone, with no music, television or other voices underneath. A phone voice memo in a quiet closet often beats a good microphone in a bare room.

Speak the way you want the voice to sound later. A calm, even read clones into calm narration; an animated read clones into energy. Use ten to fifteen seconds of continuous speech with a few different sentences, so the model hears a range of sounds, and avoid long pauses, coughs and laughter.

If your best take sits inside a longer recording, upload the whole file and set the start time to the strongest 15 seconds. A character voice keeps 10 seconds, so for a character, pick the 10 seconds that sound most like them.

• One speaker, a quiet room, no music.
• Ten to fifteen seconds of continuous speech.
• Read in the tone you want back.
• Use the start time to pick the best section.
• Keep the level steady: no clipping, no whispering.

In Seed Audio, mention the voice by its @handle, put the exact words in quotation marks, and describe the delivery and the space: "@ana-voice, warm and unhurried, close to the microphone in a quiet room, says: …". The handle brings the sound of the voice; the description shapes the performance.

Pick the length before you write. Seed Audio returns exactly the duration you choose, 5, 10, 30, 60 or 120 seconds, padding or trimming to fit, so match the words to it. People speak two to three words a second at a natural pace, which puts a 10-second line at 20 to 30 words. Speech rate, loudness and pitch are settings of their own when the read needs a nudge.

One Seed Audio run takes either up to three audio references or one image reference, not both. To score a scene from a picture, run it without the voice, and add the voice in a separate clip.

• Mention the @handle before the line.
• Put the exact words in quotation marks.
• Describe the delivery and the room.
• Choose the duration first, then fit the words.

The simplest route is a character voice. When you create or edit a character, open Character Voice and upload an MP3 or WAV (the first 10 seconds from your start time are kept), or generate one from the character’s image with an optional description, priced in Gems on the button. From then on, Cherry, Raspberry, Plum, Lemon and MiniMax-H3 use it whenever that character speaks, unless you turn Character Voices off. The AI character generator guide covers saving characters.

You can also mention an audio reference directly in a video prompt. Cherry, Plum and MiniMax-H3 take up to three, Lemon up to five and Blueberry one, and character voices count toward the same limit. Cherry needs at least one image in the prompt, such as a character or a reference, when you attach audio.

For a line that must be exact, generate it in the cloned voice with Seed Audio, then open Lipsync with a portrait or a saved character and up to 15 seconds of that audio. The AI talking photo guide shows the whole flow.

Clone only your own voice or a voice you have clear permission to use. Imitating a real person without their consent is not allowed on Mage, and the community guidelines prohibit impersonating friends, celebrities, public figures or brands for harmful purposes, and any content that is illegal where you live. If someone asks you to remove a creation that features them, remove it.

Get permission in writing, and make it specific: whose voice, what it will say, where it will appear, whether the use is commercial and for how long. A voice actor’s agreement should name synthetic voice use explicitly; a general recording release often does not cover it.

The law points the same way. In the US, the FCC ruled in February 2024 that AI-generated voices count as "artificial" under the robocall law, so using one in a robocall without consent is illegal, and Tennessee’s ELVIS Act, in force since July 2024, bars using AI to mimic a person’s voice without permission. This page is not legal advice; check the rules where you publish.

Video models’ providers screen audio too. A run can be refused because a clip contains copyrighted music or a real person’s voice, and a refused run is refunded.

• Your own voice: go ahead.
• Someone else’s voice: written permission first.
• Public figures and celebrities: not without a license.
• Never to deceive: no fake calls, endorsements or statements.
• Label realistic synthetic speech where a platform asks.

Clone a voice when the speaker has to be someone specific: you, a voice actor who agreed to it, or a character you already voiced. When any good voice will do, preset voices are faster: AI text to speech reads up to 5,000 characters in one of 16 ElevenLabs voices, and the AI voice generator covers directing a read with audio tags and describing a new voice in a prompt.

Eleven Speech does not clone voices on Mage: it speaks in its preset voices only. Cloning runs on Seed Audio and on the video models that read audio references and character voices. Developers can do all of it through the audio generation API.