Image to Text Converter

Why is the OCR making so many mistakes on my phone photo?

A few common reasons: the photo resolution is too low (characters under 20 pixels tall are hard for Tesseract), the lighting is uneven (shadows create false text edges the engine mistakes for noise), the angle is off by more than 5 degrees (the deskew pre-processing handles small angles but not heavy perspective), or the text has stylized or very small fonts that fall outside the training distribution. For best results, photograph directly overhead, fill most of the frame with the text region, use even indirect lighting without shadows, and shoot at the highest resolution your phone offers. A typical phone at 12MP shooting a letter-sized page from 30cm away produces a 300 DPI image that OCRs well.

Can this handle handwriting?

Poorly, honestly. Tesseract was trained primarily on printed text, and handwriting — with its variability in stroke, slant, connectivity, and character shapes — falls outside the training distribution. Clear block-print handwriting (each letter separated, uppercase, consistent size) works at maybe 70–80% accuracy on a good day. Cursive or casual handwriting drops to 30–50% or fails entirely. Dedicated handwriting OCR (Google Cloud Vision's handwriting mode, Microsoft Azure Read API) is significantly better because it uses models specifically trained on handwriting, but no current tool — including commercial — handles cursive reliably.

Which languages are supported?

English ships bundled. Tesseract 5 supports 100+ additional languages via downloadable trained data files, including all major European languages, Cyrillic scripts, Arabic, Hebrew, Chinese (Simplified and Traditional), Japanese, Korean, Thai, Hindi, and many more. To use a non-English language, select it from the language dropdown; the trained data file (typically 10–50MB per language) downloads on first use and caches for the session. Accuracy varies significantly — well-trained languages with large training corpora (French, Spanish, German) are near English quality; smaller languages and complex scripts are weaker.

Is my image sent to a server for OCR?

No. Tesseract.js runs entirely in your browser via WebAssembly. When you upload an image, it is decoded into a canvas, the pixel data is handed to the WASM-compiled Tesseract engine, and recognition happens locally. No network request sends any image data to any server. The only network activity during OCR is the initial download of language trained data files from a CDN (cached after first use) and standard page analytics that never see DOM or image content. This matters because people often OCR confidential documents — invoices with bank details, ID scans, internal meeting notes — where server-side OCR would be a compliance problem.

How fast is the recognition?

Depends on image size and content density. A typical letter-sized document at 300 DPI (around 2500x3300 pixels) with roughly 500 words of text takes around 4–8 seconds on a modern laptop. A single-line screenshot of a URL takes under a second. A 10-page scanned invoice PDF (processed one page at a time) takes around a minute. The first run of a session is slower because the trained data file has to download and the WASM module warms up; subsequent runs are noticeably faster. For batch processing of hundreds of documents, a server-side OCR pipeline with GPU acceleration will be 10–50x faster, but for ad-hoc work the browser version is usually sufficient.