Create speech and voice-over

Last updated: October 12, 2026

Turn a written line into spoken audio, such as a character's dialogue, a trailer voice-over or a tutorial narration. Use it when you need a voice line without booking a recording session, or a placeholder take to cut against before the final recording.

Before you start

Open the session you want the take in. Results land on that session's board.
Have the script ready, with real punctuation. Commas and full stops set the pacing.
Each run costs Creative Units. The cost shows on the Create speech button before you run. See Creative Units for how it is charged.

Steps

In the side panel to the right of the board, under Tools, click Create speech. The speech form opens in the panel. If the panel shows a tool instead of the menu, click the back arrow at the top left of the panel first.
The form opens on the default model. To try a different voice model, click the model name at the top of the form and pick another; each model has its own set of voices and controls.
Click in the prompt box (What should be said) and type the line exactly as it should be spoken, for example: The gate has held for three hundred years. Tonight, it falls. Gather your blades, and follow me into the dark. Write it in the language the line should be spoken in.
Click Advanced and describe the delivery in Style instructions: who is speaking, the mood and the pace, for example battle-weary, low and gravelly, speaking slowly. Keep the delivery out of the line itself, or the model may read it aloud.
Click Create speech, the purple button at the bottom of the panel.
Wait while the line is voiced. A tile appears on the board straight away and shows its progress; you can keep working while it runs.

Result

The take replaces the progress tile on the session's board as an audio file with its waveform. Press play on the tile to listen, or open it to see the line and the settings it was made with. To save it, right-click the tile and choose Download.

Choose a voice

Each model speaks in its own voices, so the quickest way to hear a different voice is to switch the model. Where a model offers a choice of voice or a speaker description, it appears in the form or under Advanced.

For a recurring character, keep one voice across the whole project. If your team has trained a reference set on that character's voice, click Reference set above the prompt and add it, so every line comes out in the same voice. See Create a reference set.

Direct the delivery

Models take direction in one of two ways. Some read a separate description of how to speak, in Style instructions. Others read cues written into the line itself, such as tags in square brackets or punctuation like an ellipsis for a pause. Some models offer sliders under Advanced, such as Speed or Stability. The prompting guides show what each model listens to, with lines to start from:

Production tips

One line per run. Generate dialogue a line or a short paragraph at a time. A single wrong word in a long take costs the whole take; a short one is quick to redo.
Spell names the way they are said. Invented names, numbers and acronyms are read the model's way. Write Ael-theer rather than Aelthyr if it must be said the same way across a cast, and write numbers out as words.
Keep the direction fixed across a scene. Reuse the same model, voice and Style instructions for every line a character says, and change only the words, so the lines cut together.
Localise with the same setup. For each language, run the translated line with the same model and direction. Check that the model speaks the language before you plan a full localisation pass.
Clean up and reuse. Open the take and use Change voice to move it to another voice, or Clean audio to remove noise. To put the line in a character's mouth, use it as the audio in Lip sync image or Lip sync video; see Lip sync a character to a voice line.

If something goes wrong

The model reads the direction aloud. Move descriptions of the delivery out of the line and into Style instructions, or use the cue format that model's prompting guide describes.
A word or name is mispronounced. Respell it phonetically in the line, or split the line so you can retake only that part.
The delivery is flat or wrong. Be specific about the speaker, the mood and the pace in a few plain words, and run it a few times; each run gives a different take. Try another model for a different voice.
The generation failed. The tile on the board shows the failure. Run it again; if it fails again, shorten the line or change the model. See Fix a failed or stuck generation.