Tesseract 5 Fine-Tuned to Read Mansi, Cutting OCR Errors 20-Fold

Why standard OCR fails

Mansi uses Cyrillic with extra characters: long vowels marked with a macron, such as а̄, о̄, ӯ, э̄ and ы̄, and the letter ӈ for the sound «ng». Vowel length changes meaning, so losing a macron can turn one word into another. The standard Russian Tesseract model does not know these characters. It drops the macron or turns the letter into garbage, and it misreads about 18 of every 100 characters. A second problem is that Unicode can store the same accented letter in several ways, which look identical to people but differ for a model.

Synthetic training data

Because almost no scanned Mansi books exist, the team generated its own data. Python scripts took real Mansi text from printed and electronic publications, normalized it to Unicode NFC so each letter has a single representation, and rendered it in several dozen typefaces, including FreeSerif, DejaVu Sans, Arial, Calibri and Times New Roman. Lines were drawn at several sizes and in regular, bold and italic styles.

The team also labeled each character with a bounding box. Automatic font metrics often miscalculate boxes for accented letters, so the researchers added pixel-level corrections per character, point size and typeface. The scripts produced 1.3 million pairs of line images (.tif) and ground-truth text files (.gt.txt).

The base was the official Russian model rus.traineddata from the tessdata_best repository. Since the Mansi alphabet largely overlaps with Russian, the network mainly had to learn the new macron letters and ӈ.