How Akruti stores Marathi
Most consonants are stored as a half form and completed by a bar, so म is Ù followed by ð. The short i is typed before its consonant and the reph after its syllable. क, क्र and फ are two glyphs, a body and a hook, and a vowel sign may be typed between them: ¨÷î is के. The converter reads each piece and puts the letters in Unicode order.
Why other tools see nothing here
The Akruti fonts that book embeds renumber their codes 1 to 91. Read by those codes, the text is control characters; read the way PyMuPDF and IndicPDF’s ordinary route read it, it is the accented letters in the sample. Neither is Marathi. The typed keys survive only as the names the PDF gives each glyph, and that is where this converter reads them from. Until it did, the ordinary route wrote the Latin letters into Word, and the legacy route did not notice a legacy font at all.
The colon key is a second trap. Inside a word it is the visarga: all 41 colons followed by a letter in the book are स्वतः, दुःख, निःसंशय and their like. At the end of a word it is usually a label, as in प्रकाशक: or मूल्य:, so there a Devanagari word list decides: स्वतः becomes a word with a visarga, and संकेत: keeps its colon.
Measured
On मी नास्तिक का आहे, nine pages, the Word file holds 4,930 words, 94.8% of them in Tesseract’s Marathi word list, with no sign sequence Marathi cannot spell. It agrees with a reference converter on 4,977 of 4,979 words, but that converter came with the mapping, so the agreement is not independent evidence. The two words that differ are ऐ, which the book types as ए and a sign: this reads याऐवजी where the other reads the impossible एे. A rendered page of the result reads word for word as printed. The sample is one book.
Where it still goes wrong
The table is named for the three faces that book embeds: AkrutiMar_BYogini, AkrutiMar_BAditi and AkrutiMar_BVijay. Another Akruti release may place a letter differently, and nobody has checked one. It reads PDFs only, not Word files typed in Akruti, and it cannot write Unicode back into Akruti. Text the same pages set in other fonts does not come across: the book’s years and prices in Times and its web and email addresses are left out, and so is a back-cover blurb set in an Akruti Unicode font whose map is broken.