Text to video with voiceover

One voice track, not one clip per scene

Narration is assembled into a single track laid over the scene sequence, and that track is the source of truth for timing, for the captions and for the length of the finished composition. It is not a set of clips each trapped inside its own scene. That distinction is what stops a video full of unnatural pauses where a line finished early and the cut had not arrived yet.

Dynamic stretches a little, calm stretches fully

The two pace settings resolve the clash differently. On dynamic, a scene grows by at most one and a half times to accommodate its line, and whatever is left of the line simply flows over the next cut, which is how ad editing actually works. On calm, the scene stretches to fit the whole line before cutting, which is what a tutorial needs because the viewer is following along on their own screen.

Per scene control once it exists

Every scene has its own narration text, its own language and voice, a delay in seconds before the voice starts, a playback speed of 0.85, 0.9, 1, 1.1 or 1.25 times, and a caption toggle. Regenerating narration grows the scene to fit the new audio and remembers the original duration as a base, so a shorter rewrite shrinks the scene back rather than leaving dead air at the end.

What the AI picks and what it does not

Your text becomes the script and the on-screen copy, but the model also makes editorial choices around it: which template to use, the order of the scenes, how long each one runs and which scene role carries each part of the message. Those choices can differ between runs, so the same text generated twice can come back as two different videos. What the model never touches is the machinery: layouts, motion, timing curves and rendering are fixed code. So editing a narration line afterwards retimes its scene but does not rearrange the video around it.

Which languages get a voice

The voice is chosen from the language of the brief. 56 languages have a mapped voice, most of them on Google Chirp 3 HD, with Malay, Filipino and Traditional Chinese on an older voice tier. A language with no mapping gets no voiceover at all rather than falling back to English, although the on-screen text is still written in that language. Tutorials are always given the instructional voice style, whatever the brief asks for, because a calm steady read suits a viewer following steps.

How it works, in three steps

1. Step 1

   Paste the text as a brief

   Up to 10,000 characters. Write it as what you want said, not as a description of a video, and the director will divide it into scenes.

2. Step 2

   Pick the pace deliberately

   Dynamic for anything promotional, calm for anything instructional. It changes how narration and cuts negotiate, not just how it looks.

3. Step 3

   Fix the lines that run long

   Open the voice dialog on any scene that feels rushed, shorten the line or drop the speed to 0.9, and the scene retimes itself.

The full walkthrough with screenshots is in the guide Add voiceover, captions and music to a video.

Limits worth knowing

• Narration is synthesised text to speech. There is no emphasis markup, no pause tags and no per word control beyond delay and speed.
• Captions are derived from the narration text, so editing an on-screen caption does not change what the captions say.
• Each voiceover clip you generate in the builder is billed by length, about half a credit for a typical line.
• The brief box takes 10,000 characters, which is a long script but not a document.

Questions people ask

Related pages