Methodology
How we test
Most "best handwriting OCR" lists rank tools by reputation. We rank them by measurement. Every tool is given the identical input, and its output is scored the same way against the same reference. Here is exactly how, so you can check it or repeat it.
The sample
A single handwritten page of standard English prose, 100 words, written in legible modern cursive on white unlined paper. It is deliberately a fair, middle-of-the-road case: no fade, no marginalia, no exotic vocabulary, no historical script. The goal is a baseline every serious tool should handle, so that differences reflect the engine rather than a trick page.
The reference
The page was transcribed by hand and double-checked to create the ground truth. Every tool's output is compared against this same reference text.
The input
Every tool received the same image at the same resolution: a clean 300 DPI scan, roughly 2 MB. The LLM vision models were given a strict "transcribe this exactly, do not paraphrase, do not correct" instruction. Tesseract used its default invocation.
The metric
We use Word Error Rate (WER), the standard metric in OCR and speech recognition. WER counts substituted, missing, and inserted words, divides by the number of words in the reference, and returns a percentage. Lower is better; 0% is a perfect transcription. We also note reading-order failures separately, because whole lines appearing out of sequence are far more disruptive to fix than isolated word errors, and raw WER understates them.
The measured results
| Tool | Type | WER |
|---|---|---|
| Handwriting OCR | Handwriting specialist | 0.9% WER |
| Azure Document Intelligence | Cloud document AI | 8.67% WER |
| AWS Textract | Cloud document AI | 10.5% WER |
| Claude (vision) | LLM vision | 11.2% WER |
| GPT (vision) | LLM vision | 14.4% WER |
| Google Document AI | Cloud document AI | 23.3% WER |
| Transkribus | Trainable specialist | 47.7% WER untrained |
| Tesseract | Open source | 95.4% WER |
What we are honest about
Single sample. One page is enough to surface order-of-magnitude differences, not to split hairs. A 3-point WER gap between two mid-table tools is within the noise of any single document; a ten-fold gap is not.
Legible English prose. Cursive difficulty, faded ink, historical scripts, and non-Latin languages all shift the numbers. The measured table is a baseline; the reviews discuss how each tool behaves outside it.
Phone apps are assessed, not WER-scored. Tools whose natural input is a phone photo rather than a comparable scan (Google Lens, Apple Live Text, Pen to Print) are reviewed from hands-on use and clearly marked as not measured, so nothing reads as a benchmark number when it isn't.