Why Tamil recognises comparatively well
Tamil has a smaller consonant inventory than most Indic scripts and, crucially, far less vertical stacking. Where Telugu and Kannada build conjuncts by placing a subscript form beneath the base letter, Tamil generally writes consonant clusters in sequence and marks a bare consonant with a dot above it, the pulli (்). Letters therefore stay on the line and stay separable, which is exactly what a recogniser needs. That is reflected in the measurements below.
Where it still goes wrong
Several Tamil letters differ only in a small stroke or loop, and those pairs are the usual source of substitutions at low resolution. Vowel signs that attach to both sides of a consonant (as in ெ ா combinations) can be split across a line break by a naive reader. Grantha letters used for Sanskrit-derived sounds — ஜ, ஷ, ஸ, ஹ — appear less often in training material and are correspondingly less reliable.
Measured accuracy
On our ground-truth corpus — 2 Tamil documents, 3,286 characters of prose — this pipeline reads 99.1% of characters correctly (0.88% character error rate), the strongest result of any script we measure. The corpus is clean digital renders; photographs and photocopies score lower. Two documents is a small sample, and we would rather say so than imply more coverage than we have.