Extraction or recognition
An Odia PDF saved from a Unicode document holds its text, so converting it is extraction and comes out close to exact. A scanned Odia PDF holds pictures of pages, so it is recognised first and inherits recognition’s mistakes, described on the Odia OCR page. Some Odia PDFs made by online converters hold a text layer that is itself broken, with conjuncts in the wrong order; the converter notices the damage page by page and reads those pages from their pixels instead.
Odia PDFs made by Word
Word writes one text for each glyph in a PDF, the text of the first word it drew that glyph in, so a conjunct used in several words copies wrongly in all but the first. The converter reads each glyph from the shaping rules of the font Word embedded, Nirmala UI or Kalinga, and puts the letters back in order. One limit remains: Nirmala UI draws the wa and ba subscripts with the same glyph, so ସ୍ୱ can come back as ସ୍ବ, and nothing in the glyph says which it was.
Akruti Oriya, the legacy layout
Odia DTP was typed in Akruti’s legacy faces, which store keys rather than Odia letters, so their text copies as Latin letters. The table here was read off the fonts themselves, and its InDesign variant from Akruti’s own catalog; a PDF set in either is read automatically. It has not yet been checked by a reader who reads Odia, and the converted document says so.
What is kept
Odia text, reading order and paragraphs come through, and so do the fonts a digital PDF names, with their sizes, bold, italic and colour; its tables, as Word tables; its pictures; and its running heads, in Word’s header. A PDF whose every page is scanned is rebuilt a page at a time, at the page’s size, with its lines where the scan sets them. Text wrapped round a picture, or a box laid across columns, comes out simpler than the page.