Skip to content
kura

Transcripts and search

Build a greppable spoken-word corpus from a channel's transcripts, capture transcripts only for a light archive, pick languages, and grep across an entire channel.

A YouTube archive's most valuable text is often not the description, it is the transcript: a channel's entire spoken content, made searchable on disk. kura stores every video's transcript in two shapes so the archive is both faithful and greppable.

How transcripts are stored

For each video and each captured language, kura writes two files under videos/:

  • <vid>.transcript.<lang>.vtt is the timed transcript, the source.
  • <vid>.transcript.<lang>.txt is the flat transcript text, grep-friendly.

The .vtt preserves the cue timings so the rendered watch page can show a timestamped, readable transcript block. The .txt strips the timing so a plain text search across a channel just works.

A transcripts-only corpus

For a light, fast archive of just the spoken word, --transcripts-only captures each video's transcript text and skips the heavier sidecars:

kura archive @lexfridman --transcripts-only

This is the cheapest way to make a whole channel full-text searchable. It writes the records and the transcript files but no streams, so a large channel stays small.

Picking languages

--lang selects which transcript language or languages to store. Without it, kura stores the default track:

kura archive @mkbhd --lang en
kura archive @arte --lang en,fr,de

Grepping the corpus

Because the flat transcript text is on disk, a channel's entire spoken content is searchable with the tools you already have:

grep -rl "the thing he said about" ~/data/kura/youtube/@lexfridman/videos/
grep -rn "neural network" ~/data/kura/youtube/@lexfridman/videos/*.txt

The Markdown view embeds the full transcript inline too, so grep -r over md/ searches the readable archive the same way. No other tool in the family gives you this for YouTube.

When a transcript is gated

From a flagged egress IP, YouTube serves an empty transcript body (poToken gating). kura detects the empty timedtext, records the gap in the manifest, and falls back to the optional yt-dlp path for transcript text when a yt-dlp binary is available. A clean IP gets the transcript directly. The archive is honest about exactly which transcripts it could not capture rather than silently dropping them.

Next

2 min read