Raw transcripts are mostly not content
Automatic transcription produces timestamps on every line, a speaker label on every line, and the full texture of spoken language: the ums, the you-knows, the false starts, the repeated words. A one-hour meeting transcript can be twelve thousand tokens of which a meaningful fraction carries no information.
Cleaning first routinely removes twenty to thirty percent, and the summary that comes back is usually better, because the model is not spending attention on noise.
When to keep timestamps and speakers
Strip timestamps unless you need the model to cite moments — a meeting summary that says "at 14:32 the decision was made" needs them, and a topic summary does not. They are expensive: a timestamp on every line is several tokens per line.
Keep speaker labels for anything involving attribution, disagreement or action items, since a summary that cannot say who committed to what is much less useful. Drop them for pure content extraction.
Merging turns
Transcription tools split on pauses, so one person speaking for thirty seconds becomes eight lines with eight speaker labels. Merging consecutive turns by the same speaker removes the repeated labels and gives the model continuous prose to read, which it handles better than fragments.