English | မြန်မာဘာသာ | ဘာသာမန်
Mon is spoken by roughly one million people across Myanmar and Thailand. UNESCO classifies it as vulnerable.
MonOCR takes an image of Mon script and returns text. It runs in the browser at ocr.mondevhub.com, offline once the model is cached. The Android and iOS apps build from source and are not in an app store yet.
Built and maintained by the Mon developer community.
- Web: ocr.mondevhub.com
- Android: not yet on Google Play — build from
apps/android - iOS: not yet on the App Store — build from
apps/ios
Three real documents, the text the CLI returned for each, and the per-line records
behind it — nothing hand-corrected. See samples/, which also states
what was screened out and why the headline number is not an average.
Two of the three carry non-Unicode text layers, one Zawgyi and one legacy 8-bit, and both are among the cleanest results: rasterisation happens before the model, so no encoding ever reaches it.
All three apps ship one model, and since 2026-08-15 it is v3.5:
| Architecture | MobileNetV3-Large + SE + 2×BiLSTM-512 + attention + CTC |
| Parameters | 11,553,437 |
| Input | Grayscale, 160px height, static 1024px width |
| Charset | 276 characters, 277 classes |
| Precision | FP32 |
| Published at | janakhpon/monocr, revision d3d9d5e |
Android and iOS bundle it (46.2 MB and 46.3 MB respectively). The web app fetches it from that pinned revision. Per-app details are in apps/android, apps/ios and apps/web.
v3.5 is not a newer v2, it is a different contract. Input height went 128 to
160, output classes 316 to 277, charset 315 to 276, and the graph's width axis
went from dynamic to a static 1024. Anything still holding a cached v2 artifact
is refused rather than decoded, because a mismatch of that kind returns
well-formed Mon text that is wrong. v2 remains served at revision a51be11
for anyone pinned to it.
Held-out CER 0.0100 on 150 unseen lines in a typeface the model never trained on (95% interval 0.0056 to 0.0147), measured 2026-08-16. Read it with the limits the model card states beside it:
- n = 150, so the interval is wide.
- One typeface. All 150 lines are Pyidaungsu, held out from training.
- Unseen text, not an unseen renderer. Training and test images came from the same generator, so the number says nothing yet about camera photographs of real pages.
- Disjointness is argued, not directly verified.
The figure that selected the checkpoint, val_cer 0.0210 over 4,096 lines
during training, is not an evaluation and is not comparable to it.
A v4 server model was archived on 2026-08-05. It was never trained to convergence, so archiving it was a decision about maintaining a second path rather than about measured quality. It is not maintained.
No device latency number exists for any platform. Figures of that kind appeared here until 2026-08-15 and were architectural estimates, never measured on hardware.
Because high-quality Mon datasets are scarce, validated samples from the app's feedback flow feed directly into future training rounds.
Each app runs the model in this format:
| Platform | Format | Execution provider requested |
|---|---|---|
| Web | ONNX | WebGPU where the browser offers it, otherwise WASM |
| Android | ONNX | NNAPI, with CPU fallback |
| iOS | CoreML .mlpackage |
Core ML, all compute units |
The provider is what each app asks for, not what was measured. Whether NNAPI or the Neural Engine actually runs the BiLSTM layers has not been checked on a device, and they are not expected to.
- Web App — SvelteKit PWA
- Android App — Jetpack Compose
- iOS App — SwiftUI
- Feedback Service — Go ingestion API
- Shared Assets — model weights, locales, sync scripts
- HuggingFace — ONNX and CoreML exports, and the model card
- npm package — JavaScript SDK
- Architecture decisions — ADRs
- API specs — OpenAPI contracts
- Mon Corpus Collection — training dataset
- Bugs: GitHub Issues
- Translations: Shared translation sheet
- Script samples: Contribute via the Android or iOS app, or reach out directly
- Standards: Contributing Guide · Security Policy