Bengali image to text turns printed Bangla into text you can edit, search and share, without typing it out letter by letter. It’s useful on both sides of the border, in Bangladesh and in West Bengal, for the same kinds of jobs:
- Textbook pages, guide books and question papers for notes and revision
- Notices, circulars, gazettes and application forms
- Newspaper columns and magazine articles
- Poems and quotes shared as images on Facebook
- Older books and family papers that only exist in print
Bengali and English are both selected on this page, because so many forms, notices and textbooks mix the two. Both are read in the same pass.
What makes Bangla text tricky
In Bengali, the line along the top of the letters is called the matra (মাত্রা). It joins letters into words, so there are no gaps between letters for the engine to use, and some letters carry only part of the matra or none at all, which breaks words into uneven pieces.
The vowel signs, called kar, move around the consonant. ি, ে and ৈ are written before the consonant they belong to, even though they’re typed after it. ো and ৌ wrap around it, with one part on each side: কো is ক with ে on the left and া on the right. The engine has to put all of these back in typed order.
Juktakkhor, the joined consonants, are where Bengali gets dense. ক্ষ, জ্ঞ, ঙ্গ, ন্ত, ক্ত and স্ত্র each squeeze two or three letters into one shape, and several look little like the letters inside them. Small signs carry meaning too: hasanta (্), khanda ta (ৎ), anusvara (ং), visarga (ঃ) and chandrabindu (ঁ).
The engine reads Bengali with a model trained on the script, so clear printed Bangla is usually read well. When it does go wrong, it’s usually in the kar, the juktakkhor and those small signs.
Bijoy PDFs and Unicode
A great deal of Bangla text, especially in Bangladesh, has been typed with the Bijoy keyboard in non-Unicode fonts such as SutonnyMJ. These fonts store each Bangla letter as an ordinary Latin character and only look like Bangla on screen. Copy text from a PDF made this way and you get a jumble of English letters and symbols.
Image to Text App takes a digital PDF page’s text as it is, so a Bijoy PDF shows the same jumble here. To get real Bangla, read the page as a picture: open the PDF, take a screenshot of the page, and paste it with Ctrl+V. The result is Unicode Bangla, the kind that works in Facebook, Gmail, Google Docs, Word and every modern app. Scanned PDFs are always read with OCR, so they need no workaround. See PDF to text for more on how PDF pages are handled.
Bangla, English and Assamese
Admission forms, bank forms and many textbooks switch between Bangla and English, and both are read with this page’s selection. If a page is entirely in Bangla, you can untick English in the language menu (the globe), which gives the engine fewer wrong options. The page is read again when you close the menu.
Numbers come through as printed, in Bengali digits (০১২৩) or Western ones (0123). Small digits are easy to misread, so check roll numbers, dates, amounts and phone numbers against the image.
Assamese isn’t in the language list. It uses the same script with a few different letters, such as ৰ and ৱ, so reading it as Bengali gives a draft in which those letters, and the words around them, need checking by hand. Other Indian languages have their own options, including Hindi; see Hindi image to text for Devanagari.
Old books, newspapers and question papers
Newspapers set Bangla in narrow columns. Crop to one column or one article so lines from neighboring stories don’t run together.
Older books were often printed with letterpress type, uneven ink and older forms of some juktakkhor. Expect more corrections than with modern print. Raise Contrast or try Black & white for yellowed pages, and Reduce noise for speckled photocopies.
Question papers and guide books are often photocopies of photocopies. A slightly larger, sharper copy makes more difference than any setting, because the kar and small signs are the first details to blur away.
For a whole book chapter or a stack of pages, photograph each page, drag them into order, and use Whole document view to search across them and export one file.
Getting a cleaner result
- Make the text big enough. Move closer rather than zooming, or zoom in on the screen before taking a screenshot.
- Use the original image. Images shared on Facebook and WhatsApp are compressed, and the small signs suffer first. A direct photo or the original file reads better.
- Light the page evenly and keep your shadow off it. The photo to text page has more on photographing pages.
- Straighten angled photos with four-corner straightening so the matra runs level.
Auto improve is on by default. It inverts dark-mode screenshots, enlarges small text, straightens a slight tilt and evens out lighting, and lists each step it took. For more, read how to get accurate OCR results.
Checking what it read
Words the engine was unsure about are underlined, and a button jumps from one to the next. Click any line and its place lights up on the image. Check these first: a kar attached to the wrong letter or dropped, juktakkhor read as two separate letters, a missing hasanta or chandrabindu, and pairs that differ only by a dot, such as ব and র, য and য়, or ড and ড়. If a name is misread the same way throughout, find and replace fixes every copy at once. To type corrections, turn on a Bangla keyboard in your phone or computer’s settings.
Where the text goes next
Copy the text into Google Docs, Word, Facebook or an email, and it pastes as normal Unicode Bangla. Rich copy keeps headings and lists. You can also download:
- Word (.docx) for editing and printing. See image to Word.
- Searchable PDF, the original image with an invisible Bangla text layer you can search and select
- PDF, which for Bangla opens your browser’s print dialog, where you choose “Save as PDF” so every letter is drawn correctly
- Plain text or Markdown for notes
Private by default
Certificates, forms and family papers are read in your browser, on your device, and nothing is uploaded. History is off unless you switch it on. Read more in is online OCR private?