HOW IT'S BUILT The banner is driven by word-level timestamps taken from the generated audio (forced alignment / ASR), stored alongside each clip — so the highlight tracks the real speech, not a guess. Here it's simulated.