Tesseract vs TrOCR
Pick Tesseract for printed pages and PDFs on a CPU box, in many languages. Pick TrOCR for English handwritten text lines, where it reports 2.89 CER on IAM with the large model. TrOCR only reads single text-line images, so a full page needs a separate line detector in front of it.
Tesseracttesseract/tesseract | TrOCRmicrosoft/trocr | |
|---|---|---|
| Type | Open source | Open source |
| Price | Free | 1 credit / page |
| Per 1,000 pages | $0.00 | $1.00 |
| Typical speed | 0.6 s / page | 1.2 s / page |
| Best for | Clean printed text, free | Handwritten text lines |
| Inputs | PDF, images | images |
| Languages | 100+ languages | English |
| Max pages per request | 500 | 100 |
| Tables | No | No |
| Handwriting | No | ✓ Yes |
| Math | No | No |
| Bounding boxes | No | No |
| Confidence | No | No |
| Forms | No | No |
| Multilingual | ✓ Yes | No |
| On OCRRouter | Available | Available |
Prices are what you pay on OCRRouter. Speed is the provider's typical figure, not a benchmark we ran.
Which to choose
Choose Tesseract if
- It recognizes more than 100 languages out of the box, and you can train it for more.
- It runs on CPU only with no model server, so it fits small machines and offline batch jobs.
- It outputs plain text, hOCR, searchable PDF, TSV, ALTO and PAGE XML, and it is the engine behind OCRmyPDF.
Choose TrOCR if
- It reports 2.89 character error rate on the IAM handwriting set with the large model, and Tesseract is not built for handwriting.
- It has separate checkpoints fine-tuned for handwritten (IAM), printed (SROIE) and scene text.
- It is MIT licensed, loads in a few lines with Hugging Face transformers, and can be fine-tuned on your own line images.
Key differences
| Tesseract | TrOCR | |
|---|---|---|
| Input unit | Pages and scans; returns lines and words | Single text-line images only; you need a separate line detector |
| Handwriting | Not built for it; official models target printed text | Handwritten checkpoint reports 2.89 CER on IAM (large model) |
| Languages | More than 100 out of the box, trainable for more | English-focused checkpoints; other languages need fine-tuning |
| Hardware | CPU only, no GPU needed | Base 334M and large 558M parameters; slow versus CRNN recognizers on CPU |
| Output formats | Plain text, hOCR, searchable PDF, TSV, ALTO, PAGE | Plain text |
| Licence | Apache-2.0 | MIT (microsoft/unilm repository) |
| Known failure | Accuracy drops without binarization, deskew and about 300 DPI input; confuses 1 with l | No layout, table or page-level output |
Sources
- Tesseract GitHub README
- Improving quality (tessdoc)
- Issue #2753 1 vs l
- TrOCR paper (arXiv 2109.10282)
- trocr-large-handwritten model card
What each one is
TrOCR
Transformer text recognition model. Reads single text lines, including handwriting.
TrOCR detailsTry both on your own file
Run the file once with each model and compare the output.
PDF, PNG, JPG, WebP, HEIC. Up to 20 MB each, 10 files.No file?
Use one, fall back to the other
{"model": "tesseract/tesseract","fallbacks": ["microsoft/trocr"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Tesseract vs olmOCRTesseract vs KrakenPaddleOCR (PP-OCR) vs TrOCRDocling vs TrOCRAmazon Textract vs TesseractPaddleOCR-VL vs TesseractDocling vs TesseractMinerU vs TrOCR
Tesseract alternativesTrOCR alternativesBest OCR models in 2026