timestamp_start and timestamp_end in seconds, marking where in the source video that match occurs. Precision depends on which signal produced the segment.
Frame and scene segments
Frames are extracted at fixed intervals (every 5 seconds) — aframe match’s timestamps are anchored to that grid, not to the exact sub-second moment something appeared. scene segments are bounded by detected shot boundaries, which can be tighter than 5 seconds if the video cuts quickly, but aren’t frame-perfect either.
In practice: treat a frame or scene timestamp as “within a few seconds of this point,” not a frame-accurate cue. If your use case needs sub-second seeking, seek slightly before timestamp_start in your player rather than exactly to it.
Transcript segments
transcript timestamps come from Whisper’s speech segmentation, which is generally tighter — segment boundaries follow natural pauses in speech rather than a fixed grid. This makes transcript matches more reliable for “jump to exactly where this was said” than visual matches are for “jump to exactly where this appeared.”
Using timestamps to seek video
video_url is a presigned URL to the full video file, not a pre-clipped segment — append a fragment or use your player’s API to seek to timestamp_start:
#t= media fragment for MP4. If you need a pre-trimmed clip rather than a seek point, you’ll need to trim it client-side or in your own pipeline — Pureframe AI doesn’t generate clip files.