T

sebastjan df6011c3cf Detect Scribe hallucinations + filter from SRT + auto-retry

Bug found in Žena ME TEPE third re-test:
- Scribe transcribed only verse 1 (0-33s) properly
- Then returned a single 98s segment [34.7-133.2] with just 1 word 'sam'
- This is a known Scribe hallucination on instrumental sections
- Result: SRT showed 'SAM SAM SAM SAM...' 14 times across the chorus
- Looked completely wrong because the chorus audio was correct but
  subtitles showed 'SAM' repeatedly

Three-part fix:

1. SRT GENERATOR: skip segments > 15s with < 5 words.
   These are hallucinations and have no real transcription value.

2. SCRIBE TRANSCRIBE: detect hallucinations in returned segments.
   - Mark segments > 15s with < 5 words as hallucinations
   - Compute true coverage % (excluding hallucinations)
   - Add _hallucination_count and _coverage_pct to result

3. TRANSCRIBE_FULL: auto-retry Scribe if quality is poor.
   - If hallucinations detected OR coverage < 50%, retry once
   - Keep retry result only if it has better stats
   - Otherwise fall back to first attempt (still better than nothing)

This makes the pipeline robust against Scribe's occasional bad transcripts
on songs with long instrumental breaks. Most second attempts succeed
where the first failed (random Scribe variance).

2026-04-29 18:08:35 +00:00

app

Detect Scribe hallucinations + filter from SRT + auto-retry

2026-04-29 18:08:35 +00:00

scripts

Detect Scribe hallucinations + filter from SRT + auto-retry

2026-04-29 18:08:35 +00:00

templates

Fix: live preview blocks right-side preview buttons

2026-04-29 15:37:51 +00:00

.env.example

Initial: reels clipper app

2026-04-28 15:28:22 +00:00

.gitignore

Initial: reels clipper app

2026-04-28 15:28:22 +00:00

docker-compose.yml

Initial: reels clipper app

2026-04-28 15:28:22 +00:00

Dockerfile

Add Deno runtime for yt-dlp YouTube nsig challenge solving

2026-04-28 16:05:09 +00:00

README.md

Initial: reels clipper app

2026-04-28 15:28:22 +00:00

requirements.txt

ACRCloud auto-recognition: never block uploads, fall back to fingerprinting

2026-04-29 14:24:53 +00:00

README.md

Reels Clipper · biba.live

Self-hosted Opus Clip alternativa za FOLX TV / PTC. Pretvori 16:9 video v 9:16 reels/shorts/tiktok format z auto face tracking, podnapisi (sl/de/en) in avto-detekcijo refrena v glasbenih pesmih.

Features

📤 Drag & drop upload (do 2 GB)
📺 YouTube URL paste (yt-dlp)
🎯 Smart reframe: track (face follow), center, blur (za glasbo)
🎵 Auto-chorus detection (Whisper + energy hibrid)
📝 Burned-in podnapisi (faster-whisper, multi-jezik)
🎨 3 stili podnapisov: reels, yellow (MrBeast), minimal
🔐 HTTP Basic Auth
📊 Real-time progress (Server-Sent Events)
📦 Docker / Coolify ready

Quick start (lokalno)

docker compose up --build
# odpri http://localhost:8000

Default login: sebastjan / nastavi AUTH_PASS v .env.

Coolify deploy

V Coolify ustvari nov projekt → Docker Compose iz tega repoja
Domena: reels.biba.live

Env vars:

AUTH_USER=sebastjan
AUTH_PASS=<močno geslo>
MAX_UPLOAD_MB=2000

Volume reels_data se ustvari avtomatsko
Deploy → Coolify postavi Traefik reverse proxy + SSL via Let's Encrypt

Pipeline

Upload / YouTube
       ↓
[ yt_download.py ]    ← samo če YouTube
       ↓
[ find_chorus.py ]    ← samo če auto_chorus=true (Whisper + RMS analiza)
       ↓
[ reframe.py ]        ← 16:9 → 9:16 (track / center / blur)
       ↓
[ subtitle.py ]       ← Whisper transkripcija + burn-in
       ↓
   reel.mp4

API

POST /api/upload — multipart file upload, vrne job_id
POST /api/youtube — JSON {url, mode, lang, ...}
POST /api/process — start processing za uploaded job
GET /api/jobs — list vseh
GET /api/jobs/{id} — status
GET /api/stream/{id} — SSE stream progress
GET /api/download/{id} — final reel
DELETE /api/jobs/{id} — pobriši

Dependencies

FFmpeg (system)
faster-whisper (transkripcija)
OpenCV (face detection)
yt-dlp (YouTube)
FastAPI + uvicorn (server)