Drop scanned PDFs or photos (JPG/PNG/WebP) and get editable text back. Tesseract — the open-source OCR engine — runs as WebAssembly inside this tab; the language files (English ~4 MB, Simplified Chinese ~2.5 MB, the speed-optimized "fast" variants) download from this site only when that language is first used.
Every page reports a confidence percentage, and anything under 70% carries a ⚠ check-manually flag. PDFs are rasterized page by page at 2× before recognition; images are capped at a sane upscale so a 4000-pixel phone photo stays a 4000-pixel canvas.
Frequently asked questions
How accurate is it?
On clean 300-DPI scans of plain text, very good. On phone photos with perspective skew, low contrast, handwriting or complex multi-column layouts — disappointing. The per-page confidence number tells you which pages to trust; treat anything below 70% as a draft.
Why is Chinese slow?
The fast Chinese model still evaluates far more character classes per glyph than English letters, and this is the wasm build on CPU — expect 10-30 s per page. The engine swap planned for a future version (PP-OCR via ONNX Runtime WebGPU) would cut that several-fold.
Does it OCR handwriting?
Not usefully — Tesseract is a print-recognition engine. Handwriting needs dedicated models that do not exist in a browser-viable form yet.
中文说明
拖入扫描版 PDF 或照片（JPG/PNG/WebP），取回可编辑文字。开源 OCR 引擎 Tesseract 以 WebAssembly 在本标签页内运行；语言文件（英语约 4MB、简体中文约 2.5MB，速度优化的 fast 版）只在首次选用该语言时从本站下载。
每页报告置信度百分比，低于 70% 附 ⚠ 人工核对标记。PDF 逐页以 2× 栅格化后再识别；图片限制合理放大倍数，4000 像素的手机照片不会变成超大画布。
常见问题
识别准确率如何？
清晰 300 DPI 的纯文本扫描件：很好。带透视倾斜、低对比度、手写体或复杂分栏的拍照件：会失望。逐页置信度数字告诉你哪些页可信；低于 70% 请当作草稿。
为什么中文慢？
fast 版中文模型每个字形仍需评估远多于英文字母的字符类别，且这是 CPU 上的 wasm 构建——每页约 10-30 秒。未来版本计划的引擎更换（PP-OCR + ONNX Runtime WebGPU）可提速数倍。
能手写识别吗？
不实用——Tesseract 是印刷体识别引擎。手写需要专用模型，目前尚无浏览器可用的形态。