hamohand/tessdata-kab

OCR tamazight kabyle # Tesseract OCR for Kabyle / Amazigh (Latin script) This repository contains the `kab.traineddata` language model for Tesseract OCR (v4.0.0 and v5+), dedicated to the **Kabyle language (Taqbaylit)** and standard **Amazigh** written in the Latin alphabet. This model is a community contribution developed as part of the **Dawaliw-ⵣ-OCR** project. ## 🌟 Model Overview - **Language Code:** `kab` (ISO 639-3) - **Engine:** LSTM only - **Base model:** Fine-tuned from `fra` (French) - **Supported script:** Latin (with extended Amazigh characters) - **Current Version:** v6.0 (August 2026) ## 🔤 Supported Characters The model recognizes the 33 standard Amazigh letters, including the specific diacritics used in Kabyle linguistics, as well as their uppercase equivalents: * **Base Latin:** a, b, c, d, e, f, g, h, i, k, l, m, n, q, r, s, t, u, w, x, y, z * **Extended Amazigh (Special Characters):** * `ɣ` (Gamma) * `ḥ` (h with dot below) * `ɛ` (Open E) * `ḍ`, `ẓ`, `ṭ`, `ṣ`, `ṛ` (Emphatic consonants with dot below) * `č`, `ǧ` (Caron) * `ţ` (t with cedilla - specific to Kabyle) * `ʷ`, `ᵒ` (Modifier letters for labiovelarization) *(Note: The model correctly normalizes the Greek epsilon `ε` to the Latin open E `ɛ` to respect strict Unicode standards for Amazigh data).* ## 📊 Performance (v6) The model was trained on both synthetic data (78k unique Kabyle words) and fine-tuned on real historical scans (e.g., *Dictionnaire Kabyle-Français J.M. Dallet*). - **BCER (Character Error Rate):** 1.50% - **BWER (Word Error Rate):** 3.42% - On recent, clean, single-column documents, accuracy on special Amazigh characters reaches **95% - 100%**. ## 📥 Installation 1. Download the `kab.traineddata` file. 2. Place it in your Tesseract `tessdata` directory. - **Linux:** `/usr/share/tesseract-ocr/4.00/tessdata/` or `/usr/share/tesseract-ocr/5/tessdata/` - **Windows:** `C:\Program Files\Tesseract-OCR\tessdata\` - **macOS (Homebrew):** `/opt/homebrew/share/tessdata/` or `/usr/local/share/tessdat …