Voice, timing, and revisions

The voice-over is not a soundtrack laid under the animation — it is the clock the animation runs on. Get one take with word-level timestamps and every cue in the video anchors to a word, so the number lands on the number and the title card lands on "three".

One take. Always.

One TTS call for the entire script. Not one per scene, not one per step.

Per-scene narration produces an audible seam at every join, a flat list-like delivery, and a timing model that drifts the moment any scene changes length. One take also gives you one continuous word table, which is what makes cue-chaining possible.

Request word-level timestamps (/with-timestamps on ElevenLabs). Without them you are placing cues by stopwatch, and every script edit re-times the whole video by hand.