✏️ Text Input
🎙️ Microphone Input
Idle

🔒 On-device Whisper (transformers.js). First run downloads the model (~100–200MB). No audio leaves your device.

🎬 Video Upload → Sign
No file selected

Extracts audio via ffmpeg, transcribes with Whisper, then plans and renders sign animation.

🗄️ Sign Database

Load multiple files one-by-one. Later entries override earlier duplicates by gloss.

Stored in localStorage only. Used as an optional GPT-4o-mini planner fallback.

📚 Loaded Vocabulary
⚙️ EDU NLP Pipeline

EDU NLP extracts pronouns, verb lemmas, and key nouns. Abbreviations collapse multi-word forms (e.g., "carbon dioxide" → "CO₂").

📋 Planner Diagnostics
CWASA loading…
Sign animations rendered via CWASA using SiGML + HamNoSys gloss sequences. Unmatched words are fingerspelled when the alphabet database is loaded.