A demo video's captions used to be timed by estimate: each sentence got its share of the narration by length, and the marked word moved on by its letters, so in pauses and fast stretches the caption ran ahead of the voice or fell behind it. Now the render hears its own narration back with speech recognition and times every sentence and every word from it. The caption starts each sentence with its first spoken word and marks each word the moment it is heard, for every voice. Older demos and translated subtitles keep the estimate.
Demo captions follow the voice word by word
Shared by Daniel Sadilek on November 9, 2026