Skip to main content
Restrict a query to spoken words with modes=transcript. This searches transcribed speech only — visual content, scene labels, and on-screen text are excluded from matching.

​
How it works

Every video’s audio is transcribed by Whisper into timestamped speech segments. Transcript search runs Postgres full-text search (websearch_to_tsquery) against that text — a lexical match, not a vector embedding. This means it rewards queries that share actual words with what was said: "pricing objection" matches a segment containing those words, but won’t match a segment that expresses the same idea in different words the way visual/conceptual search can. See Embeddings for why the two techniques are different under the hood.

​
When to restrict to transcript-only

  • You know the moment involved someone saying something specific, and want to avoid visual false-positives
  • You’re building a “search what was said” experience (meeting notes, lecture search, call review) where visual matches aren’t relevant
  • You want faster, more precise results on content where speech is the primary signal — talking-head interviews, voiceovers, panel discussions
For everything else, the default combined mode (see Combined Queries) is usually the better choice — it catches moments whether they’re identifiable by what’s shown, what’s said, or both.