Original research
Handwriting OCR benchmark: Tesseract, cloud OCR and AI tested on 183 handwritten pages
8 OCR and handwritten text recognition tools, 183 public handwritten pages, one scorer. Every answer key, every output and the code are open.
Short answer: free OCR engines can't read handwritten notes. Tesseract found 1 handwritten word in 10, and the best free model, TrOCR, about 2 in 3. The cloud OCR services from Microsoft, Google and Amazon read 84 to 87% but still missed 11 to 14% of the words. An AI model (ScanPilot) scored 90.9% and missed 7%. On printed forms with handwritten fill-ins, every cloud service and ScanPilot scored within 0.6 points of each other.
Results
- AI model
- Cloud OCR
- Open-source OCR
| ScanPilot | |
|---|---|
| Azure Document Intelligence | |
| Google Cloud Vision | |
| Amazon Textract | |
| TrOCR (large, handwritten) | |
| PaddleOCR (PP-OCRv5) | |
| EasyOCR | |
| Tesseract 5 |
| System | Group | All pages183 pages | Handwritten notes172 pages | Forms11 pages | Clean pages101 pages |
|---|---|---|---|---|---|
| ScanPilot | AI model | 91.6% | 90.9% | 94.6% | 94.5% |
| Azure Document Intelligence | Cloud OCR | 88.0% | 86.6% | 94.5% | 92.9% |
| Google Cloud Vision | Cloud OCR | 86.2% | 84.5% | 94.1% | 91.9% |
| Amazon Textract | Cloud OCR | 86.0% | 84.1% | 94.7% | 90.8% |
| TrOCR (large, handwritten) | Open-source OCR | 67.0% | 64.3% | 79.5% | 72.1% |
| PaddleOCR (PP-OCRv5) | Open-source OCR | 54.4% | 48.9% | 79.2% | 65.1% |
| Tesseract 5 | Open-source OCR | 24.4% | 13.1% | 68.3% | 36.0% |
| EasyOCR | Open-source OCR | 23.1% | 15.0% | 61.0% | 33.0% |
Word F1 in percent; the best score in each column is in bold. Recall is the share of the handwritten words in the answer key that a tool found. Precision is the share of the words it wrote that are really on the page. F1 combines the two, so a tool scores high only if it finds most of the words and invents few. "Clean pages" are the pages with nothing crossed out or unreadable, where the answer key is most complete. Recall, precision and results by writer region are in the GitHub repo.
What we found
- General OCR engines do not read handwritten notes.
Tesseract found 10.5% of the handwritten words and EasyOCR 15.2%. Both do much better on forms (68% and 61%), where most of the text is printed.
- A model trained on handwriting makes the biggest difference among free tools.
TrOCR, built for handwritten lines, scored 64.3% on notes, 15 points above PaddleOCR and about 50 above Tesseract and EasyOCR.
- The three cloud OCR services are close to each other.
Azure led Google and Amazon in every writer region, by 1 to 4 points. All three still missed 11 to 14 of every 100 handwritten words.
- Forms are close to solved, notes are not.
When handwriting sits in printed boxes, the cloud services and ScanPilot all scored 94.1 to 94.7%. The differences show on free-form notes.
- The AI model missed the fewest words.
ScanPilot missed 7 of every 100 handwritten words, more than a third fewer than Azure (7.1% against 11.3%). It led in all four writer regions (Africa, Asia, Europe, North America). It is our product and the only AI model in this test; see the disclosure below.
How accurate is handwriting to text?
- AI model
- Cloud OCR
- Open-source OCR
| ScanPilot | |
|---|---|
| Azure Document Intelligence | |
| Google Cloud Vision | |
| Amazon Textract | |
| TrOCR (large, handwritten) | |
| PaddleOCR (PP-OCRv5) | |
| EasyOCR | |
| Tesseract 5 |
So even the best cloud OCR leaves about one word in nine to fix by hand on a page of notes.
Can Tesseract read handwriting?
Not well. Tesseract is the free, open-source OCR engine behind many apps and most Python tutorials, and it is built for printed text. On our handwritten notes it found about 1 word in 10 (13.1% F1). Its sparse-text mode (--psm 11) did worse. The open-source alternatives:
- EasyOCR: about the same as Tesseract on notes (15.0%).
- PaddleOCR: reads about half the words (48.9%). Its recogniser is documented as supporting handwriting.
- TrOCR: the best free option (64.3%). It reads one line at a time, so it needs a text detector in front of it; we used PaddleOCR's.
What OCR can and cannot do with handwriting in general: OCR for handwriting.
Tesseract vs Google Vision vs Azure vs Textract on handwriting
On handwritten notes, every cloud service read more than 6 times as much as Tesseract: Azure 86.6%, Google Cloud Vision 84.5%, Amazon Textract 84.1%, Tesseract 13.1%. On forms the gap is smaller (about 94% against 68%) because most form text is printed.
Google Vision vs Azure vs Textract: which cloud OCR reads handwriting best?
Azure Document Intelligence (the Read model) scored highest on notes and needs no handwriting setting. Google Cloud Vision was tested with its handwriting option on, which added about 1.5 points. Amazon Textract was tested with plain text detection and is not offered in every AWS Region. On forms the three were within 0.6 points, so for form filling any of them will do.
AI OCR vs traditional OCR
ScanPilot reads the page with an AI model, with its own page handling and layout step around it. On notes it scored 4.3 points above the best traditional cloud OCR and missed more than a third fewer words. On forms it was level with them. Two things to know:
- It sometimes fixes the writer's spelling ("memtorship" becomes "mentorship", about 1 word in 80). The answer keys keep the original spelling, so those count as misses here.
- It failed on one page, a student's handwritten copy of a published text, which its model declined to transcribe.
We did not test general AI models on their own (such as ChatGPT or Claude), so this test does not say how they compare.
To see what an AI model does with your own handwriting, upload a page to the handwriting to text converter; the first page is free, no signup. For a tool-by-tool overview, see the best AI OCR software.
Handwriting recognition in Python: which library to use
Handwritten text recognition in Python comes down to four choices:
- pytesseract (Tesseract) and EasyOCR: fine for printed text, not for handwriting (13 to 15% on our notes).
- PaddleOCR: about half the words. Its default models need a lot of memory on large phone photos; the smaller "mobile" detector worked better for us.
- TrOCR (Hugging Face transformers): the best free choice for handwriting (64.3%).
- A cloud API: if you need more than two thirds of the words, call Azure, Google or Amazon (84 to 87%), or an AI model. The runners for every tool in this test are in the GitHub repo.
TrOCR reads one cropped line of handwriting at a time:
from PIL import Image
from transformers import TrOCRProcessor, VisionEncoderDecoderModel
name = "microsoft/trocr-large-handwritten"
processor = TrOCRProcessor.from_pretrained(name)
model = VisionEncoderDecoderModel.from_pretrained(name)
line = Image.open("line.jpg").convert("RGB") # one cropped line of handwriting
ids = model.generate(processor(images=line, return_tensors="pt").pixel_values)
print(processor.batch_decode(ids, skip_special_tokens=True)[0])
How we tested
- Pages
- 172 phone photos of handwritten notes from the GNHK dataset (ICDAR 2021, CC BY 4.0), written in Europe (82 pages), North America (36), Africa (29) and Asia (25), and 11 scanned forms with handwritten fill-ins from the OmniAI OCR benchmark (MIT). The answer keys hold 9,199 handwritten words.
- Answer keys
- Written by the dataset authors, not by us and not by any tool tested.
- Same input
- Each photo turned upright first. The cloud services received identical files, and the open-source tools the same upright image.
- Settings
- Default settings, plus each provider's handwriting option where one exists. The open-source tools ran out of the box, without extra training. ScanPilot was run three times because AI output varies slightly; we report the average (on the notes, the three runs were within 0.4 points). Every other system was run once. Exact versions and settings are in the table below.
- Scoring
- Word-level comparison with the answer key, ignoring word order, case and punctuation. Full rules, every output for every page and the scorer are in the GitHub repo, so anyone can check or rerun the test.
- Dates
- 3 November 2026 (ScanPilot, Google, Amazon, Tesseract) and 5 November 2026 (Azure, EasyOCR, PaddleOCR, TrOCR). Services change; we will rerun and publish a new version when they do.
- Speed
- Not compared. The open-source tools ran on one laptop and the cloud times depend on network and region, so timings would not be fair.
| System | Version and settings | Run on | Runs |
|---|---|---|---|
| ScanPilot | Build of 3 November 2026, layout mode, default settings | 3 Nov 2026 | 3 |
| Azure Document Intelligence | prebuilt-read model, API version 2024-11-30 (v4.0). No handwriting setting exists | 5 Nov 2026 | 1 |
| Google Cloud Vision | DOCUMENT_TEXT_DETECTION with the handwriting language hint (en-t-i0-handwrit) | 3 Nov 2026 | 1 |
| Amazon Textract | DetectDocumentText (plain text detection), Region eu-central-1 | 3 Nov 2026 | 1 |
| Tesseract 5.5.1 | English LSTM model (--oem 1), automatic layout (--psm 3) | 3 Nov 2026 | 1 |
| EasyOCR 1.7.2 | English reader with default settings | 5 Nov 2026 | 1 |
| PaddleOCR 3.4.1 (PP-OCRv5) | PP-OCRv5 mobile detector and server recogniser, orientation and unwarping off | 5 Nov 2026 | 1 |
| TrOCR (large, handwritten) | microsoft/trocr-large-handwritten on text lines found by the PaddleOCR detector | 5 Nov 2026 | 1 |
Limitations
- English only for now. Hindi, Bengali, German, Japanese, Chinese and other languages are next.
- The GNHK answer keys keep the writers' spelling mistakes, which penalises tools that correct spelling.
- 11 forms is a small sample, so the forms row shows "all close", not a ranking.
- One AI model. General AI models on their own and phone apps such as Google Lens, Microsoft OneNote and Adobe Scan were not tested.
- The open-source tools could do better after training on handwriting. This test measures what they do out of the box.
- The test pages are public. GNHK has been online since 2021 and the OmniAI pages since 2025, so they may be in the training data of any AI model, including the one behind ScanPilot. We cannot rule that out from the outside. A check on handwritten pages that have never been published is planned for the next version.
- PaddleOCR's recogniser is multilingual and wrote 0.8% of its words in Chinese, Japanese or Korean characters. They count as wrong words, as they would for a user.
Disclosure
This study was run and written by Luigi Paolo Ceci, founder of ScanPilot and managing director of SUSTRAVIA GmbH, the company that publishes it. ScanPilot is our product, so we have an interest in the result. That is why the data, the answer keys, every tool's raw output and the scoring code are public. If you think a tool was set up unfairly, tell us the setting you would use (open an issue in the GitHub repo or write to hello@sustravia.com) and we will rerun it and publish the result, whichever way it goes.
Frequently asked questions
Can OCR read handwriting?
It depends on the tool. On our handwritten notes, free OCR engines like Tesseract missed most words, cloud OCR services missed about 1 word in 8, and ScanPilot about 1 in 14.
What is the best handwriting OCR?
In this test, ScanPilot scored highest on handwritten notes (90.9% word F1), followed by Azure Document Intelligence (86.6%). On forms with handwritten fill-ins, the leading tools were level.
Is Tesseract good for handwriting?
No. Tesseract is designed for printed text and found about one word in ten on our handwritten notes.
What is the best free handwriting OCR?
TrOCR (microsoft/trocr-large-handwritten), at 64.3% on notes. It reads one line at a time, so you need a text detector to split the page into lines first.
Is AI OCR better than traditional OCR for handwriting?
In this test, the AI model missed more than a third fewer words on notes than the best traditional cloud OCR. On forms there was no difference.
How do I convert handwritten notes to text?
Take a clear, upright photo with even light, then use a handwriting OCR tool. With ScanPilot you upload the photo and download the text or a table as Excel or CSV.
Can I use this data?
Yes. The answer keys, outputs and results are open (CC BY 4.0, code MIT). Please credit "ScanPilot handwriting OCR benchmark" and the original dataset authors.
Cite this study
ScanPilot (2026). Handwriting OCR benchmark: Tesseract, cloud OCR and AI tested on 183 handwritten pages (v1.0). Zenodo. https://doi.org/10.5281/zenodo.24002850
Data and code: github.com/ScanPilot/handwriting-ocr-benchmark. DOI: 10.5281/zenodo.24002850. Changelog: 1.0 (5 November 2026), English set, 8 systems.