Phonetic Support Architecture for WHG

• 1. Overview
• 2. Components
  • 2.1. Online Components (DigitalOcean)
  • 2.2. Offline Components (Pitt CRC)
  • 2.3. Rationale for Separate Indices
• 3. Data Flow
  • 3.1. Initial Migration (One-Time, at Pitt)
  • 3.2. Ongoing Ingestion (New Datasets)
  • 3.3. Linking Toponyms to IPA
  • 3.4. Linking IPA to Places (Implicit)
  • 3.5. IPA Generation: Epitran + PanPhon
  • 3.6. IPA Normalisation Rules
• 4. Elasticsearch Index Design
  • 4.1. IPA Index
  • 4.2. Toponym Index
  • 4.3. Place Index
• 5. Training the Siamese BiLSTM
  • 5.1. Training Data Construction
  • 5.2. Model Architecture
  • 5.3. Embedding Refresh Cycle
  • 5.4. Real-Time Inference Model
• 6. Query Pipeline
  • 6.1. Full Pipeline with Fallbacks
  • 6.2. Query-Time Optimisations
  • 6.3. Error Handling
• 7. Push-Based Synchronisation Strategy
  • 7.1. Rationale
  • 7.2. Bulk Update Workflow
  • 7.3. Authentication & Security
  • 7.4. Resilience Strategy
  • 7.5. Verification Strategy
• 8. Advantages of This Architecture
• 9. Monitoring & Observability
  • 9.1. Key Metrics
  • 9.2. Dashboards
  • 9.3. Alerting
• 10. Future Extensions
• 11. Deployment Plan
  • 11.1. Phase 1: Development (Week 1-4)
  • 11.2. Phase 2: Initial Migration (Week 5-8)
  • 11.3. Phase 3: Model Training (Week 9-12)
  • 11.4. Phase 4: Production Rollout (Week 13-16)
  • 11.5. Phase 5: Continuous Improvement (Ongoing)
• 12. Risk Assessment
• 13. Success Criteria
  • 13.1. Technical Metrics
  • 13.2. User Experience Metrics
  • 13.3. Research Impact
• 14. Summary
• 15. References