README.md

metadata

title: شَمْل - Arabic Diacritization
emoji: 🎙️
colorFrom: indigo
colorTo: purple
sdk: docker
app_port: 7860
pinned: false
license: apache-2.0
short_description: Arabic speech-to-text + diacritization

شَمْل (Lisan / Shaml) — Arabic Diacritization

AI-powered Arabic diacritization platform combining:

• Whisper ASR for Classical Arabic speech-to-text
• Diac (NAACL 2024) transformer for restoring tashkeel
• FastAPI backend serving a React (Vite) frontend
• Optional Supabase persistence for audio + text records

Workflow

1. Record or upload an Arabic audio clip
2. Whisper transcribes it
3. Edit the transcription if needed
4. Diac restores diacritics on the edited text
5. (Optional) Save the audio + texts to Supabase

Environment Variables (Secrets)

Set the following as Repository Secrets in the Space settings (not Variables):

• SUPABASE_URL — your Supabase project URL
• SUPABASE_SERVICE_ROLE_KEY — backend-only secret (never expose client-side)

Optional:

• API_KEY — protect the API behind a header (X-API-Key)
• MODEL_NAME — override default Hugging Face Diac model
• ASR_MODEL_NAME — override default Whisper model
• CORS_ORIGINS — comma-separated allowed origins (default permissive)

Resources

• Free tier: 2 vCPU / 16 GB RAM / 50 GB disk (non-persistent)
• First boot downloads Whisper + Diac models from Hugging Face (~3-4 GB)
• Subsequent boots use cached models

API

• GET /api/ — service info
• GET /api/health — model & Supabase status
• POST /api/transcribe/audio — Whisper transcription
• POST /api/diacritize — text diacritization
• POST /api/diacritize/audio — one-shot audio → diacritized text
• POST /api/records/save — save to Supabase
• GET /docs — interactive OpenAPI

Credits

Based on the NAACL 2024 paper:

Shatnawi, S., Alqahtani, S., Aldarmaki, H. (2024). Automatic Restoration of Diacritics for Speech Data Sets.