Timestamps are the part of a transcript people use constantly without thinking about. They're also the source of several behaviours that look like bugs and aren't. The model underneath is simple enough to explain in a paragraph.
What a segment contains
A transcript is a list of segments. Each one holds exactly three things:
- The text — the words in that chunk
- A start time — how many seconds into the video the chunk begins
- A duration — how long it stays on screen
That's the whole model. Times are stored in seconds and displayed as [mm:ss] or [h:mm:ss] for longer videos. Everything a transcript does with time follows from these three fields.
Why lines break oddly
Segment boundaries are chosen so captions appear on screen readably — roughly the amount of text someone can take in during a couple of seconds. They have nothing to do with grammar.
So a long sentence gets split across several segments, and a fast speaker may fit an entire exchange into one. This is the direct cause of the chopped-up look of raw transcripts, and it's why punctuation is missing too — both come from the same design decision.
How click-to-seek works
Clicking a transcript line tells the player to jump to that segment's start time. That's all it is — the line knows what second it belongs to, and the player is told to go there.
The same data drives highlighting in the other direction: as the video plays, the current time is compared against segment start times to work out which line is active. That's why the transcript scrolls along with playback.
Why timing feels slightly off
Clicking a line sometimes lands you a beat early or late. Several ordinary reasons:
- Captions lead the audio. Text usually appears slightly before the words are spoken so viewers have time to read it.
- Segments overlap. Consecutive captions can overlap slightly to avoid a visible flicker between lines.
- Paragraph view merges segments. A merged block takes the start time of its first segment, so clicking it jumps to the beginning of the block rather than the exact sentence you clicked near.
- Player buffering. Seeking has to load video at the target point, and the player may settle on the nearest available keyframe.
None of these are errors. Caption timing is built for comfortable reading, not frame-accurate synchronisation.
When to keep them and when to drop them
Keep timestamps when you're navigating the video, citing a specific moment, or building notes you'll need to verify later. Drop them when you're reading the transcript as prose, pasting it into a document, or feeding it to another tool — there they're pure noise, and in the last case they consume tokens without adding anything.