playtranslate/ocr-models

PlayTranslate OCR models

On-device OCR model packs for the PlayTranslate Android app, converted to MNN format for mobile inference. Each subdirectory is one downloadable pack referenced by the app's model catalog; the app fetches the individual files directly and verifies a pinned SHA-256 per file.

Packs

Pack	Files	Size	Source	License
meiki-ja	det.mnn, rec_horizontal.mnn, rec_vertical.mnn	46.7 MB	Meiki OCR — D-FINE detector + horizontal/vertical recognizers (Japanese)	LGPL-3.0
paddle-rec-cjk	rec.mnn, keys.txt	16.6 MB	PaddleOCR PP-OCRv5 — Chinese (Simplified/Traditional) + Japanese + English	Apache-2.0
paddle-rec-latin	rec.mnn, keys.txt	8.0 MB	PaddleOCR PP-OCRv5 — Latin scripts (French, Spanish, German, …)	Apache-2.0
paddle-rec-korean	rec.mnn, keys.txt	13.4 MB	PaddleOCR PP-OCRv5 — Korean	Apache-2.0
paddle-rec-arabic †	rec.mnn, keys.txt	8.0 MB	PaddleOCR PP-OCRv5 — Arabic	Apache-2.0
paddle-rec-cyrillic †	rec.mnn, keys.txt	8.0 MB	PaddleOCR PP-OCRv5 — Cyrillic (Russian, Ukrainian, …)	Apache-2.0
paddle-rec-devanagari †	rec.mnn, keys.txt	7.9 MB	PaddleOCR PP-OCRv5 — Devanagari (Hindi, Marathi, …)	Apache-2.0
paddle-rec-thai †	rec.mnn, keys.txt	7.9 MB	PaddleOCR PP-OCRv5 — Thai	Apache-2.0
manga-ocr-ja	encoder.mnn, decoder.mnn, vocab.txt	67.2 MB	manga-ocr-base-2025 — ViT-tiny + 2-layer decoder whole-block Japanese recognizer (opt-in MangaOCR refinement)	Apache-2.0

† Hosted ahead of in-app language wiring (support coming soon). The first four packs are live in the app; these four are downloadable and integrity-pinned but not yet mapped to a selectable source language.

The shared PaddleOCR PP-OCRv5 DBNet text detector is bundled inside the app APK (not hosted here) and is paired with whichever recognizer pack above the source language selects.

Licenses & attribution

• Meiki OCR — GNU LGPL-3.0. Converted to MNN for on-device use; conversion scripts available on request.
• PaddleOCR PP-OCRv5 recognizers — © PaddlePaddle Authors, Apache-2.0 (https://github.com/PaddlePaddle/PaddleOCR). Converted to MNN.
• MangaOCR (manga-ocr-ja) — manga-ocr-base-2025 © jzhang533, Apache-2.0 (https://huggingface.co/jzhang533/manga-ocr-base-2025, weights at revision 1e64d5be; redistribution expressly confirmed by the author in the repo's discussion #1), based on manga-ocr by Maciej Budyś (kha-white), Apache-2.0 (https://github.com/kha-white/manga-ocr). Modification: converted to fp16 MNN for on-device use. The manga-ocr model family is trained using the Manga109-s dataset, whose use is hereby acknowledged per its terms (http://www.manga109.org/).
• MNN runtime/format — © Alibaba, Apache-2.0 (https://github.com/alibaba/MNN).