Loading
fix(transcripts): prefer timed formats, and read word-by-word JSON as sentences
An episode may list several transcripts; the parser kept the first. On
Buzzsprout that is the HTML one, shown as a single block that does not
follow the playback, while a timed JSON sits right after it. The parser
now keeps a format whose lines carry times (Podcasting 2.0 JSON, WebVTT,
SubRip, by type or extension), the first among equals; HTML and plain
text only when nothing else is offered. A feed refresh rewrites the
episode, so stored ones switch too.
That JSON is timed word by word, though: 31,729 segments for one Podnews
episode, which would have made a line per word. Such transcripts (ten
segments or more, mostly single words) are now regrouped into sentences:
a new line at the end of a sentence, when the speaker changes, after a
two-second pause, or past about 280 characters. The Podnews episode reads
as 1,139 timed lines. Transcripts already cut into sentences are kept as
they are.
Co-Authored-By:
Claude Opus 5.5 <noreply@anthropic.com>