Modern OCR narrows gap on doctors’ handwriting

A new benchmark of handwritten medical prescriptions shows that compact modern optical character recognition systems can decipher doctors’ handwriting far better than established OCR engines, while remaining well short of the accuracy needed for autonomous clinical use.

The April 2026 test compared PP-OCRv5, GLM-OCR, Tesseract and EasyOCR on cropped handwritten prescription words. GLM-OCR produced the best character-level performance, recording a Character Error Rate, or CER, of 0.328. PP-OCRv5 delivered the lower Word Error Rate, or WER, at 0.789, compared with 0.801 for GLM-OCR. Lower scores indicate fewer errors under both measures.

Exact-word recognition showed a much sharper divide between the newer and older technologies. GLM-OCR correctly reproduced 32.6% of tested words, while PP-OCRv5 achieved 21.4%. Tesseract managed 2.5% and EasyOCR 2.6%, giving the two newer systems roughly an eightfold to 13-fold advantage depending on the comparison.

The results underline an important shift in OCR development. PP-OCRv5 uses only about five million parameters, yet its word-level performance remained competitive with GLM-OCR, a 900-million-parameter multimodal model. Research behind PP-OCRv5 argues that carefully selected, accurately labelled and diverse training data can allow specialised lightweight OCR systems to compete with far larger vision-language architectures.

GLM-OCR represents the alternative approach. Its architecture combines a roughly 400-million-parameter visual encoder with a 500-million-parameter language decoder. The model is designed to interpret document images while using linguistic context to improve transcription. Its multi-token prediction mechanism also allows several tokens to be generated during a decoding step, reducing some of the speed penalty normally associated with generative vision-language systems.

That language capability appears particularly useful when individual letters are difficult to distinguish. In the prescription benchmark, GLM-OCR’s CER was considerably better than PP-OCRv5’s 0.477. Yet PP-OCRv5 performed marginally better on WER, suggesting that stronger character-level reconstruction does not automatically translate into fewer complete-word errors.

Generative OCR also introduces different failure modes. During testing, GLM-OCR sometimes produced formatting or characters that were not present in the handwritten input. One four-letter sample was interpreted as mathematical notation rather than ordinary text. PP-OCRv5, based on a more conventional specialised recognition pipeline, was less prone to this type of generative deviation, although it still misread large numbers of difficult words.

The experiment used RxHandBD, a prescription handwriting dataset containing 5,578 cropped word images and 1,559 unique text entries. The material includes generic medicine names, pharmaceutical brands, dosage forms and clinical terminology drawn from physical prescriptions. A prepared version divides the dataset into 4,463 training images and 1,115 testing images.

Those characteristics also limit how broadly the findings can be applied. The benchmark used cropped individual words rather than complete prescription pages, meaning it did not fully test the challenge of locating handwriting, separating printed material and preserving relationships between drug names, doses and instructions. It was also a single CPU-only run using default model configurations without medical fine-tuning.

Handwritten prescription recognition remains difficult because medical documents combine highly variable writing, abbreviations, unfamiliar drug terminology and contextual information that can alter the meaning of otherwise similar text. Research into multimodal prescription interpretation has highlighted the additional difficulty created by limited datasets and the loss of contextual information when prescriptions are divided into isolated word images.

The clinical threshold is considerably higher than the benchmark results. Even GLM-OCR’s 32.6% exact-match score means roughly two words out of three were not reproduced perfectly. A system making errors at that rate cannot safely determine medicines, strengths or dosing instructions without human verification. The benchmark itself was an independent technical test rather than a clinical validation study and did not evaluate patient outcomes or dispensing safety.