Why Kannada is tall
Kannada is a sister script of Telugu and builds clusters the same way: the second consonant is written as an ottakshara, a small subscript form beneath the base letter. Vowel signs sit beside or above the consonant, and some, such as ೊ and ೋ, are drawn in several parts. A single syllable can therefore be two or three layers tall. The subscript is the smallest and lowest part of it, and the first to be lost on a blurred or low-resolution scan. Losing it rarely produces obvious garbage. It usually produces a different, real Kannada word.
The anusvara that reads as a zero
Kannada draws the anusvara ಂ and the digit zero ೦ almost identically. On the live service, all three anusvaras on one Kannada page came back as zeros, ಬೆಂಗಳೂರು as ಬೆ೦ಗಳೂರು, at confidences of 95, 93 and 88, so confidence gave no warning at all. The repair is a rule of spelling rather than a list of words: Kannada does not put a digit inside a word, so a zero attached to a Kannada letter is turned back into the anusvara. A number standing on its own is left alone.
Measured accuracy
On our ground-truth corpus — 1 Kannada document, 2,176 characters of prose — this pipeline reads 95.1% of characters correctly (4.87% character error rate). That is one clean digital render, which is a small sample, and printed or photocopied pages will do worse. The test file also lists ottaksharas one after another out of context; that part scores far lower and is reported separately, because nobody writes Kannada that way.
What still gives it trouble
Deep ottakshara stacks at small point sizes, faded photocopies where the subscripts have broken up, display typefaces, and lines where Kannada and English alternate word by word. Handwriting is not supported; this reads printed Kannada. A rescan at 300 dpi or higher helps more than anything else.