How Bengali OCR Handles যুক্তবর্ণ and Handwriting

Scanning a Bengali exam script is harder than it looks. Bengali orthography uses over 300 conjunct characters (যুক্তবর্ণ) formed by combining base characters with hasanta. A single misread ligature can turn a correct answer into gibberish. E-Porikkha's OCR pipeline is built specifically to handle this.

When you upload a scan, the file is split into individual pages and each page is sent to a vision-language model with a Bengali-aware prompt. The model reads the image holistically — not character by character — which makes it far more robust to smudging, uneven ink pressure, and margin notes.

Pages are processed in parallel, so a 10-page script takes roughly the same time as a 2-page one.

Wherever the model cannot confidently read a word — due to extreme smudging, torn paper, or very faint pencil — it outputs [illegible] rather than guessing. This is intentional. A hallucinated Bengali word looks plausible but is wrong; [illegible] is honest and lets the teacher decide.

You'll see [illegible] markers in the extracted text in the grading studio. You can click any marker and manually type the correct word before approving the paper.

Many science and maths papers mix Bengali prose with English technical terms, formulas, or numbers. The OCR model handles this natively — it preserves English words in their original script without transliteration.

Scan at 300 DPI or higher. JPEG compression at low quality settings blurs fine strokes. If your scanner exports PDF, keep it at the original resolution rather than re-saving as a low-quality image.

If a paper is consistently getting [illegible] in the same spot, check whether the physical page has a fold or shadow — repositioning the paper on the scanner almost always helps.