Handwriting OCR
Handwriting OCR on ExactRead: compare vision models on handwritten notes, convert handwriting to text, accept the best output, and save the winner for similar documents.
Upload a document and compare models right here — no need to leave this page.
Handwriting to text is the hardest OCR task: quality swings with the writer, the pen, the paper, the photo angle, and the language. No single engine wins on every hand, which is exactly why ExactRead runs handwriting OCR through more than one model at once. Reading two or three outputs side by side gives a reviewer a far stronger signal than trusting one model's guess, and the confidence and warning fields flag where a model was unsure.
Before running handwriting OCR, set the document type to handwriting and add a language hint (auto, en, zh, ja, and more). Telling the workbench that a page is handwritten, and in which language, tightens the extraction prompt and helps on mixed-language notes, cursive, and small annotations.
Vision-language models such as GPT-5.4, Gemini 3.1 Pro, and Qwen3-VL tend to read handwriting to text better than plain OCR engines because they use context to resolve ambiguous strokes. Google Cloud Vision also offers dedicated handwriting detection. Compare a couple of these on your own sample — the right choice depends on your handwriting, not a spec sheet.
When a model handles a particular handwriting style well, accept its result and save it as a preference. Future uploads of the same kind of note start from that route, so you spend less time re-testing and more time reviewing. Accepted handwriting OCR output exports as TXT or JSON like any other job.
Legibility drives handwriting OCR more than any single model: cursive, faint pencil, crossed-out words, cramped margins, and skewed photos all reduce accuracy. A clear, well-lit, straight-on photo at decent resolution gives every engine a better chance. For records that matter — signatures, amounts, medical or legal notes — treat handwriting to text as a first pass to be verified, not a final answer, and lean on compare mode plus the confidence and warning fields to flag the risky lines.
Drop a handwritten document onto the workbench or click to browse. Supported inputs are PNG, JPEG, and WebP images, plus PDF for PDF-capable models.
After upload, the model list is filtered to the OCR engines that actually support your file, so you never start a job that cannot run.
Choose a single model when speed and cost matter, or compare mode to run several OCR models on the same handwritten document at once.
Each result keeps its model, status, confidence, and warnings, with the recommended output highlighted so you can judge accuracy quickly.
Accept the output that reads your handwritten document most faithfully, then copy it or export TXT or JSON for the next step.
Capability, format support, and credits come straight from the model catalog — a starting point, not a ranking. Test on your own documents to decide.
| Model | Provider | Type | Formats | Credits / doc |
|---|---|---|---|---|
| GPT-5.4 | OpenAI | Vision-language model | Image files | 3 |
| Gemini 3.1 Pro | Vision-language model | Image files | 3 | |
| Qwen3-VL 32B | Alibaba Qwen | Vision-language model | Image files | 2 |
| Google Cloud Vision OCR | Native OCR engine | Images and PDF | 2 |
Supported formats
Handwritten photos, scans, and mixed-language images
Model guidance
Compare at least two vision models for handwriting
Credit note
Saved preferences reduce future trial runs
No. Handwriting is the hardest case for any OCR. ExactRead helps by exposing competing outputs, confidence, and warnings, but important handwritten content should always be reviewed by a person.
Yes. The workbench supports language hints such as auto, en, zh, and ja, and a handwriting content type. Both sharpen the extraction prompt for handwritten and mixed-language pages.
Vision-language models (GPT-5.4, Gemini 3.1 Pro, Qwen3-VL) usually lead on handwriting, and Google Cloud Vision offers handwriting detection. The only fair test is your own sample — compare a few here.
Vision models can read cursive and mixed-language handwriting, but accuracy drops as legibility drops. Setting the handwriting type and a language hint helps; always review the output for cursive text.
The free plan includes 100 OCR credits every month, granted on sign-in. Each model run costs that model's per-document credits, and compare mode charges each selected model.