Why a Shree-Lipi PDF copies out wrong
Shree-Lipi’s 8-bit faces, SHREE-DEV7 and its relatives, put Devanagari on the slots of an English font, as Kruti Dev does. The PDF stores the keys the typist pressed, and its own map from codes to letters says those keys, so a copy of म्हणून is a run of Latin letters and symbols. The page looks right only because the PDF carries the Shree-Lipi font that draws those codes.
The SHREE-DEVX faces that InDesign embeds are the other half. Their map sends every Devanagari glyph to a private-use character, a code no font on your computer draws, so a copy comes out as boxes. The converter reads neither map. For the 8-bit faces it reads the codes through a table drawn from the fonts embedded in real documents; for SHREE-DEVX it reads each glyph by its number, which every SHREE-DEVX face shares.
What comes back in the Word file
Unicode Devanagari set in Noto Sans Devanagari, so it opens on any computer without Shree-Lipi. The document is rebuilt page by page: one Word section for each printed page with its size and margins, paragraphs with their alignment and type sizes, bold, pictures on their pages, and a running head in the Word header. English the PDF sets in an ordinary font comes across as it is. A word processor still breaks lines its own way, so a page can end a line or two from where the PDF’s did. The file opens with a note that the mapping is unverified, because it is.
Measured
Twelve issues of two Marathi monthly magazines, 636 pages set in SHREE-DEV7 faces. Converted, they hold 160,505 words of five letters or more, and 90.7% of them are in Tesseract’s Marathi word list, between 87.5% and 92.8% issue by issue. Eight of the issues were used while the table was being built; the four added afterwards, which it had never seen, read between 89.3% and 91.2%. One printed line of every issue was checked by eye against the conversion, and all twelve matched.
The sign sequences the checker flags are nearly all the magazines’ own spelling: 235 of 242 are अॅ, the way Marathi writes an English a, as in अॅप and व्हॉट्सअॅप, typed as अ and the candra sign. The rest are four stray nukta keys and three slips.
Where it still goes wrong
Thirty-three codes in those twelve issues draw glyphs nobody has named yet, and they come through as the replacement character, �, so they are easy to find. All twelve issues come from one publisher, so another publisher’s Shree-Lipi build may type some letters on other keys. The SHREE-DEVX reading is measured on one real document, a government annual report: 76% of its long words are in the Devanagari list and no code is left unnamed, which is a good sign from one report and not yet proof.
Shree-Lipi’s Gujarati, Bengali and other faces have no table. Some Hindi magazine issues set most of their text in AkrutiDevBhaskar, which no table reads yet. A scanned PDF has no codes to read and needs OCR.