clip studio
Upload the service. Review the moments.
The pipeline transcribes every word, skips the welcome, announcements, worship, and offering, and proposes the strongest complete moments. You pick which ones become captioned vertical clips — nothing publishes without you.
rundown
- Word-level transcription on the GPU
- Welcome, announcements & worship skipped
- Several complete 20–90 second moments
- You choose what gets rendered
What happens on the server
- Fetching: paste a video URL and the server downloads it (when this deployment's network allows), or upload the file directly.
- Word-level transcription (faster-whisper large-v3 with a Bible vocabulary prompt) — skipped when captions come with the URL or you attach an SRT/VTT.
- Service-structure detection: welcome, announcements, worship, offering, and prayer segments are excluded from selection.
- Clip selection: a two-pass LLM rubric (hook, emotional peak, standalone completeness, quote density) proposes several complete moments, with a deterministic fallback.
- Your review: pick which moments to render. Quality gates hold every clip to 20–90 seconds ending on a finished sentence, snapped to natural pauses.
- Rendering: centered 9:16 crop with bold lower-third captions (max 2 lines), H.264/AAC MP4.
Processing runs on this deployment's hardware; a 55-minute sermon takes about 5–6 minutes to review-ready on the reference GPU. Uploads are limited to 100 MB, and finished jobs are deleted automatically after about 24 hours. Only process recordings you own or are licensed to publish — and review every clip before it goes out.