Supercut Studio

A local SvelteKit app for browsing transcript/video buckets and rendering supercuts. It can stitch phrase or regex matches, synthesize fake sentences, songify clips, transcribe missing media with xAI STT, and browse/delete finished renders.

Requirements

• Node 22+
• ffmpeg on PATH
• aubiopitch on PATH for /songify (brew install aubio)
• XAI_API_KEY for /transcribe and /supercut AI regex suggestions
• One or more local media buckets under ./videos by default
• Source videos downloaded into those buckets, for example with yt-dlp

Local .env files are loaded by Vite/SvelteKit. On boot, the app also tries to hydrate missing env vars from ../syntax-transcripts-2026/.env; override that lookup with SYNTAX_REPO_DIR=/abs/path if needed.

Media Layout

The app treats each direct child of the buckets root as one bucket. A valid bucket has a videos/ folder; transcripts/ and supercuts/ are created or read when needed.

videos/                         # bucket root
  syntax/                       # bucket name
    videos/                     # source .mp4/.mkv/.mov/.webm/.avi and .mp3
    transcripts/                # word-level .json transcripts
    supercuts/                  # generated .mp4 renders
  another-channel/
    videos/
    transcripts/
    supercuts/

Set SYNTAX_BUCKETS_DIR=/abs/path to use a different bucket root. The songify pitch cache lives beside the bucket root as .pitch-cache.json.

Download source media into a bucket's videos/ folder before transcribing or rendering. yt-dlp works well for pulling YouTube videos or playlists into the expected local layout.

Run

npm install
npm run dev
# Visit http://localhost:5180

For production-style serving:

npm run build
node build

Useful checks:

npm run check

What's Here

• / - Library dashboard grouped by bucket. Shows video, mp3, transcript, size, word count, duration, per-row transcribe actions, and delete actions.
• /clips - Grid of rendered .mp4 files from every bucket's supercuts/ folder, with filtering, playback, download, and delete.
• /supercut - Phrase, regex, or AI-assisted regex supercut builder. Supports bucket selection, tile mode, clip limits, dry-run duration estimates, and debug text overlay.
• /sentence - Picks one clip per word and stitches a sentence they never said. Supports bucket selection, fixed seeds, morphology fallbacks, and debug overlay.
• /songify - Detects per-clip pitch with aubiopitch and arranges clips into a melody. Supports bucket selection, single-clip mode, and optional pitch shifting.
• /transcribe - Transcribes one file or all missing files in selected buckets through xAI speech-to-text. Extracts .mp3 first when needed.
• /stats - Top words and n-grams across all bucket transcripts.

Architecture Notes

• Routes live under src/routes; shared UI lives in src/lib.
• Bucket discovery and path safety live in src/lib/server/paths.ts.
• Pipelines live under src/lib/server/pipelines and accept an onLog callback for streaming progress.
• Jobs are tracked in an in-memory Map via src/lib/server/jobs.ts. Restarting the server clears job history; finished outputs stay on disk.
• The UI polls /api/jobs/[id] while a job is running and can cancel jobs with DELETE /api/jobs/[id].
• Media is served through /api/media/[bucket]/[kind]/[name] with Range header support so <video> can seek without re-downloading.
• Library file deletion uses /api/library/[bucket]/[slug]; rendered clip deletion uses /api/clips/[bucket]/[name].
• Supercut, sentence, and songify renders write into the first selected bucket's supercuts/ folder. Multi-bucket outputs include mixed in the filename.

Style Notes

• Global design tokens and reusable classes live in src/app.css.
• Palette: --bg white, --ink near-black, single --accent: #D0FE03.
• Unbounded is the main typeface; JetBrains Mono is used for logs and code-like text.
• Sections separate with whitespace or 1px rules. Status indicators use outlined or filled squares, not colored pills.
• The log panel is the primary inverted surface: black background with accent text.