runningcostabrava/tachelhit-drills

# Tachelhit Drills — language documentation & learning platform A community platform for **collecting Atlas Tamazight (Tachelhit) from people and the internet, and learning it** — built around Catalan as the reference language, with Arabic and Latin romanization alongside Tifinagh (ⵜⵉⴼⵉⵏⴰⵖ). ## What it does **Capture** — turn any content into learning material: - Paste a **YouTube / Instagram / TikTok / podcast URL** (or upload a video/audio file): the server downloads it, uses platform subtitles when available and Whisper ASR otherwise, and returns timestamped segments. - **Song mode**: paste the lyrics and they are aligned onto the ASR timestamps (ASR hears sung vocals poorly but times them well). - **OCR**: burned-in subtitles (Latin + Arabic scripts) read off video frames by a dedicated Space. - **⚡ Auto mode**: one tap runs the whole chain server-side — transcribe → correct → translate (Catalan · Arabic · Tifinagh) → create drills with clipped media → optionally render a vertical reel per drill. Survives closing the page. **Learn** — evidence-based practice: - **🧠 Spaced repetition** (SM-2): per-user review queues; the home screen shows what's due. - **Recall-first cards**: only the Catalan prompt shows/plays until you reveal; then native audio plays twice. - **🎤 Pronunciation feedback**: record yourself; your speech is transcribed and fuzzy-scored against the drill's text. - **Tests** with hints, timers, and scoring; demo videos and shorts generated by a MoviePy Space. **Document** — a growing corpus, not a pile: - Every drill carries **variety, region, speaker, provenance (source URL), license, and Latin romanization** alongside the three scripts. - **👤 Accounts** (personal token, hashed at rest): isolated SRS schedules and contributor attribution. - **📖 Corpus page**: search across all scripts, filter by variety, listen, and **verify** contributions (the review queue). - **Export**: `/corpus/export` (JSON/CSV with full provenance) and `/corpus/flywheel? …