"A language dies every two weeks. When it goes, it takes a unique way of understanding time, kinship, ecology, and the human condition — knowledge that cannot be recovered." — UNESCO Atlas of the World's Languages in Danger
Language extinction is not inevitable. It is the result of colonization, marginalization, and — until now — the absence of technology designed for communities with small speaker populations. Standard speech AI requires thousands of hours of training data. Endangered languages have minutes.
- 40% of the world's ~7,000 languages are at risk of disappearing (UNESCO, 2023)
- A language dies every 14 days — at this rate, 90% of existing languages will be gone by 2100 (UNESCO)
- Less than 4% of the world's languages have any digital presence whatsoever (Endangered Languages Project, Google)
- Few-shot AI — models that learn from 10-50 samples rather than 10,000 hours — is the only viable technical path for most endangered languages
This project does not just archive languages. It helps communities actively speak them.
flowchart TD
A[Community Elders\nSpeakers / Learners] --> B[Recording Interface\nMobile + Web + Offline]
B --> C[Audio Ingestion\nWAV / MP3 / OGG\nNoise reduction + normalization]
C --> D[Few-Shot ASR Training\nWhisper fine-tune\n10-200 sample minimum]
D --> E[(Language Archive\nSupabase + S3\nCommunity-controlled)]
E --> F[Speech Recognition\nReal-time transcription\nin endangered language]
F --> G[Lexicon Builder\nWord / phrase / sentence\nwith community annotation]
G --> H[Learning Module\nInteractive lessons for\nnext-generation speakers]
H --> I[Community Dashboard\nSpeaker count, health score,\ncontribution tracker]
I --> J[Export Tools\nPDF phrasebooks, audio packs,\nresearch-ready corpora]
E --> K[Research API\nOptional: share with\nlinguistic institutions]
| Feature | Description | Approach |
|---|---|---|
| Few-Shot ASR | Speech recognition trained from as few as 10-200 audio samples | Whisper + LoRA fine-tune |
| Community Recording Studio | Mobile-first interface for elders to record words, phrases, stories | Offline-capable |
| Lexicon Builder | Community-annotated word database with audio, IPA, context, and cultural notes | Community-owned |
| Interlinear Text | Side-by-side indigenous text + phonetic transcription + translation | Linguist-standard format |
| Next-Gen Learning | Interactive lessons built from community recordings; designed for children | Game-based + audio |
| Language Health Score | Speaker population tracking, age distribution, vitality assessment | UNESCO EGIDS scale |
| Export + Archive | Research-ready corpora, PDF phrasebooks, offline audio packs | ELAR / AILLA compatible |
| Community Sovereignty | All data is owned and controlled by the speaker community, not the platform | Self-hosted option |
| Citation | Finding | Relevance |
|---|---|---|
| UNESCO (2023) | 3,000+ languages endangered; acceleration linked to urbanization + cultural pressure | Scale of urgency |
| Baevski et al. / Meta AI (2022) | wav2vec 2.0 enables ASR with < 10 minutes of labeled speech | Few-shot architecture |
| Mager et al. (2023) | Indigenous NLP requires community co-design, not extraction | Community sovereignty model |
| Ethnologue (2023) | Only 23 languages have >50M speakers; 6,500+ have fewer than 1M | Long-tail focus |
git clone https://github.com/OliWoods-Org/foundation-living-tongue.git
cd foundation-living-tongue
npm install
npm run dev- Runtime: Node.js + TypeScript
- Speech AI: OpenAI Whisper (few-shot fine-tuning via LoRA)
- Database: Supabase (PostgreSQL + Storage)
- AI: Claude API (NLP, annotation assistance)
- Offline: Service Workers + IndexedDB
- Export: PDF generation, ELAR-compatible archive format
We actively seek contributions from indigenous community members, field linguists, computational linguists, and anyone with experience in language documentation. Data sovereignty is a core principle — no language data enters this system without explicit community consent.
- Fork the repo
- Create a feature branch (
git checkout -b feat/amazing-feature) - Commit your changes
- Push and open a PR
AGPL-3.0 — Free to use, modify, and distribute.
Built by the OliWoods Foundation
Free forever. Open source. Because every language that dies takes a world with it.