Measured accuracy
On our ground-truth corpus the pipeline reads 92.7% of the characters of its one Odia test page correctly (7.32% character error rate), 3,020 characters of prose: a children’s story, set clean in one column. With Malayalam’s it is one of the two lowest figures of any language we measure, and one page is a small sample, so read it as a direction rather than a rate. A photographed or worn page will do worse.
The two mistakes that make up most of it
Two Odia shapes account for most of the errors on that page. The i-sign on ଥ is drawn as a low hook, and the recognition model reads it as the u-sign: ଥିଲା came back as ଥୁଲା. Of 19 ଥି on the page, 3 came back right. And the ra-phala, the ୍ର under a conjunct, is often dropped: the story’s heroine ଚନ୍ଦ୍ରିକା came back right 3 times of 10, most often as ଚନ୍ଦିକା. Both errors leave a word that looks almost right, so read a result for them before you use it.
What is repaired, and what is not
One Odia confusion is repaired after reading: ବ read where the page has ୱ, in the conjuncts that use it after ସ, ଶ, ଜ, ଦ and ତ, and nowhere else, because ବ is also a common correct letter. The two mistakes above are not repaired, because the wrong form is sometimes a real word and a rule that changed it would change correct text too. A page scanned upside down is turned over before it is read, but on Auto detect an upside-down Odia page can be taken for another script first; choosing Odia avoids that. Handwriting is not supported; this reads printed Odia.