Train a custom voice

Last updated: October 15, 2026

Clone a voice from a recorded sample, or design one from a description, so every line a character speaks comes out in the same voice.

A trained voice lives in a reference set, next to the character's art, so the voice and the look travel together. Train it once and every speech generation that uses the set reads its script in that voice.

Before you start

  • Training needs the Train styles permission in your workspace role.
  • To clone, put one or more clean recordings of the voice in a reference set: one speaker, no music, no background noise. Voice models differ in how many clips they use and how long a clip can be; see Trainable models. If the set doesn't exist yet, create a reference set first, or open Clone voice from the side panel's tools, which starts a new one.
  • To design a voice instead, you need no clips, only a description of it.
  • Training costs Creative Units; the price shows on the Train voice button. See Creative Units.

Steps

  1. Open the reference set that holds your sample. On a session board, double-click its tile.

  2. In the Audio section, select Design voice. The setup opens on a voice model under Base model; open the picker to choose a different one.

  3. Under Method, choose Clone to copy the voice from your clips, or Design to build one from a Voice description (tone, age, accent, delivery). Some voice models ask for more: a Reference transcript of the exact words in your clip, which improves the clone, or Clip to clone when the set holds several clips and the model clones from one.

  4. Select Train voice. The voice appears under Audio marked Training. It trains on its own, so you can keep working; you get a notification when it is done.

  5. When the row reads Ready to use, open it and select the Finetunes tab. Play the audition clip to hear the voice before you use it.

  6. Back on the board, open Create speech, pick the voice model your voice trained on, then select Reference set and add your set. It lands in the prompt box as a chip.

  7. Type the script exactly as it should be spoken and select Create speech.

Result

The line lands on the board as an audio file, read in your trained voice. The voice itself stays in the reference set: under Audio in its Finetunes column, with every version on the Finetunes tab. Anyone in the workspace who can use the set can voice lines with it.

Production tips

  • Record the sample for the job. A clone copies what it hears: the room, the microphone and the delivery. Record in a quiet room, at the energy level the character usually speaks at, and leave out music and effects. A minute or two of varied sentences gives a steadier voice than one dramatic line.
  • Keep one voice per set, next to its character. Put the voice in the character's own reference set, so a single set carries both the look and the voice into every generation.
  • Design a voice you can't record. For a creature, a narrator you haven't cast yet or a placeholder for scratch audio, Design gives you a consistent voice from a description. Describe it the way you would brief an actor: age, timbre, accent, pace.
  • Write direction outside the script. The prompt is what gets said. Where the model offers delivery controls under Advanced in Create speech, put direction there, as Create speech and voice-over describes.
  • Retrain without breaking lines in production. Retraining adds a new version; the current one keeps working until the new one is ready, and you can switch back from the Finetunes tab.
  • Carry the voice into video. Use the finished line to lip sync a character to a voice line.

If something goes wrong

  • Train voice is greyed out. Hover it for the reason: the set has no clip to clone yet, a clip is still processing, the model clones from one clip and you haven't chosen which, or Design has no description. Design on some models takes no reference clips at all; remove the audio from the set or choose Clone.
  • Training failed with a message about the sample's length. Your clip is longer or shorter than the voice model accepts. Trim it to a few minutes of clean speech, or use a longer clip, and train again. Limits per model are in Trainable models.
  • Training failed because a sample was flagged. The provider's content policy rejected one of the clips. Remove it and train again. For anything else, see Training failed.
  • Your voice doesn't appear in the Create speech picker. The picker shows sets compatible with the model in the form. Pick the voice model your voice trained on first, then add the set.

Related