Accuracy
Scraped captions are not a transcript
Auto-captions were built to scroll under a video, not to be read as text. Pull them out and the seams show: no punctuation, no speakers, and the same words repeated three times over.
What auto-captions are actually for
YouTube's automatic captions are a live, rolling overlay. Each line is drawn on screen, then replaced by the next one a beat later - a display format, not a document format.
Scrape that feed and you get exactly what it was designed to produce: fragments timed to a video player, not sentences timed to a reader.
The rolling-caption problem
Because captions scroll, the same words often appear across two or three consecutive lines as the display catches up - one phrase spoken once, written down three times. Pull the raw feed and you inherit that duplication.
Stack that with no punctuation and no speaker turns, and a scraped transcript reads like a wall of repeated, run-on fragments - technically the right words, practically hard to use.
Transcription starts from the audio, not the captions
Real transcription does not scrape what YouTube displayed. It runs the audio through an ASR model and builds segmented text from scratch, with sentence boundaries and speaker turns based on what was actually said.
That is a structurally different output: clean lines instead of a scrolling overlay dumped into a file.
How Vocateca handles both cases
Vocateca can transcribe YouTube audio itself, on-device, rather than only pulling captions. When it does read captions, it dedupes the rolling overlay first - collapsing the repeated lines back into the single line that was actually spoken - before rebuilding the text.
Either way you get one clean, timestamped, per-segment line, not three overlapping copies of the same sentence.