OCR in the Browser: Tesseract, WASM and Why Preprocessing Matters

Retyping a quote from a screenshot is nobody's favorite activity. The OCR extractor runs Tesseract — compiled to WebAssembly — inside your browser and hands back editable text with a confidence score. No upload, no queue, no account.

The engine, self-hosted

The tool ships its own copy of everything: the tesseract worker, SIMD and fallback WASM cores, and the fast English model (~13 MB total, downloaded once and then browser-cached — after that it works offline). Recognition runs in a worker so the page stays responsive, and a lazy singleton means the expensive engine init happens once per visit, not once per image.

Preprocessing is half the accuracy

Raw photos punish OCR: uneven lighting, colored backgrounds, low contrast. Before recognition the tool converts the image to grayscale and stretches contrast to the full range — dark pixels darker, light pixels lighter, histogram spread wide. On thermal-printed receipts and phone photos this routinely turns garbled output into clean text. Sharp, straight-on, high-contrast input remains the recipe; no preprocessing rescues motion blur.

Reading the confidence score

Tesseract reports how certain it is, 0–100. Above ~80 the text is usually trustworthy verbatim; below that, expect the odd l/1 and O/0 swap — check the output against the preview before quoting it. The result pane keeps both side by side for exactly this.

Honest limits

• Handwriting and display scripts stay unreliable — that's true of all OCR, not just WASM builds.
• Layout is roughly preserved (blank-line collapsing keeps paragraph breaks) but complex multi-column layouts can interleave.
• Languages: English is bundled; other models are technically droppable-in later, not yet exposed in the UI.

Screenshots of code and documents are the sweet spot. Need the pages out of a PDF first? The PDF to Image Converter rasterizes them; want the extracted text read aloud? Pipe it through Text to Speech.