Behind the cuts
How it works
No black box and no external model API. Everything below runs on one server, in order, from a file on disk.
The pipeline
- 01
Ingest
ffprobe records the source facts. Audio is extracted to 16 kHz mono PCM, a 250 ms loudness envelope is computed, and silences are detected so cuts land in pauses rather than mid-word.
- 02
Transcribe
faster-whisper (
distil-large-v3, int8, CPU) with word-level timestamps and VAD. Outputs JSON, SRT, WebVTT and plain text. - 03
Score
Sentences are rebuilt into complete thoughts, then every 18–75 second window is scored on six signals: hook, quotability, specificity, delivery energy, clean cut points and self-containment. Penalties for filler and dangling conjunctions.
- 04
Deduplicate and diversify
Overlapping windows are dropped, and a keyword-overlap penalty stops three clips on the same sentence. The top set is ranked by score.
- 05
Render
One accurate cut, then three deliveries: 1080×1920 with a blurred bed and a quote card, 1080×1080, and 1920×1080. Captions are word-highlighted ASS, burned in; SRT and VTT ship alongside.
- 06
Publish and keep provenance
Catalog rows keep the source episode, exact in/out points, transcript and scoring breakdown. Human edits (titles, boundaries, status) survive a re-run.
The length rule
You asked for a cap that keeps at least half the clips at or below it. The pipeline takes the median of the selected clips and uses it as the cap, so by construction half the set sits at or under it — not an arbitrary number picked in advance.
Target 45 s, hard ceiling 75 s. The analyzer trims trailing sentences until a clip fits, and refuses to cut a clip below two sentences.
What is honest here
Scores are deterministic and reproducible from the transcript; they are not popularity or engagement predictions, so the site does not pretend they are. The only counters shown are real counts from the catalog and aggregate page views — no IP addresses, no cookies on public pages, no third-party scripts.
Machine, not editor
The transcript is unedited machine output and will get names and scientific vocabulary wrong. Clip boundaries are suggestions that a human can override in the admin console; that is why every clip has an in/out point you can see and a transcript you can check.
Use the data
Everything on the site has a JSON endpoint:
/api/clips— every published clip with render URLs/api/clips/<slug>— one clip with quote and timings/api/videos— episodes/api/videos/<id>/transcript— segments as JSON, or?format=srt|vtt|txt/api/search?q=— full-text search across clips and transcripts/feed.xml— RSS of the latest clips