English | မြန်မာဘာသာ | ဘာသာမန်
MonOCR is an open-source technical framework dedicated to the digital preservation of the Mon language (mnw). Classified by UNESCO as a vulnerable script, Mon lacks standardized inclusion in global OCR toolchains.
This project establishes a zero-leak privacy foundation for character recognition, enabling offline digitization of historical and community-sourced manuscripts.
The current inference engine (~6.6M parameters) is a V1 implementation optimized for low-latency edge execution. Given the historical scarcity of high-quality Mon-Burmese datasets, this platform acts as a data acquisition terminal. The integrated Feedback Service enables the collection and auditing of community-sourced manuscripts, which will directly inform the training of future, higher-capacity recognition models.
- Web: ocr.mondevhub.com
- Android: Google Play Store
- iOS: Apple App Store (Review Pending)
MonOCR maintains absolute architectural parity across all targets. While the underlying mathematical model is unified, it is delivered via platform-optimized serialization to maximize hardware-accelerated performance:
- Web/Android: Standardized via universal ONNX weights.
- iOS: Optimized for Apple Neural Engine via CoreML (
.mlpackage).
| Concern | Principal Implementation | Architectural Rationale |
|---|---|---|
| Model (Web/Android) | apps/android/.../monocr.onnx |
Deterministic cross-platform benchmarks |
| Model (iOS) | apps/ios/.../monocr.mlpackage |
ANE-optimized hardware utilization |
| Asset Sync | shared/locales/sync.mjs |
Multi-target linguistic idempotency |
| Ingestion Auth | internal/auth/middleware.go |
Perimeter security for asset ingestion |
| Native Execution | engine/MonOcrEngine.swift |
Hardware-bound inference logic |
| Attribute | Specification | Rationale |
|---|---|---|
| Model Architecture | MobileNetV3 + BiLSTM + CTC | Optimal accuracy-to-latency ratio for edge inference |
| Parameter Count | ~6.6M | Balanced for browser-bound execution limits |
| Asset footprint | ~25MB (FP32) | Optimized for delivery via edge CDNs |
| Inference Precision | FP32 / ANE-Optimized | Maximizing character fidelity in low-resource contexts |
- Web App: Browser-bound SvelteKit PWA.
- Android App: Native Jetpack Compose (NNAPI).
- iOS App: Native SwiftUI (Apple Neural Engine).
- Feedback Service: Mobile-focused ingestion API (Go).
- Core Assets: Shared assets and synchronization logic.
All technical documentation, architectural decisions, and setup guides are centralized in the Documentation Hub.
- Architecture (ADRs): Logical decision records.
- API Specifications: OpenAPI contracts.
- HuggingFace Models - ONNX, CoreML, CKPT: Core inference assets.
- NPM Package: Portable SDK.
Janakh Pon • Oung Seik Nyan • MonDevHub
- Feedback: Report technical bugs via GitHub Issues.
- Linguistic Assets: Audit our shared translation sheet.
- Dataset Acquisition: Contribute script samples via our Android or iOS applications or reach out directly.
- Technical Standards: Review our Contributing Guide and Security Policy.
Note
The Mon language is classified as a "vulnerable" language in UNESCO's Atlas of the World’s Languages in Danger.