Invoice Processing in 5 Languages: A Real-World Benchmark

Invoices are the most OCR'd document in business — and the most varied. Layouts differ per country, languages differ, currencies differ, and line items are formatted differently everywhere. To know where we stand, we benchmarked our parser on a corpus of 10,000 invoices across English, Spanish, Portuguese, Japanese, and Korean.

Field-level extraction accuracy

• Invoice number, date, and due date: 98.1% field-level accuracy across all five languages.
• Line-item tables: 96.4% of rows extracted with matching columns (description, qty, unit price, amount).
• Totals: 99.0% of invoice totals recognized without arithmetic errors.
• Currency and locale formatting preserved per language (€, ¥, ₩, $…).

Why language matters less than layout

The interesting finding: accuracy differences between languages were smaller than differences between layout families. A clean single-page invoice extracts well in any language; a dense multi-column invoice is hard in every language. This confirms the VLM approach — understanding structure is the bottleneck, not the alphabet.

CrazyOCR detects the invoice language automatically, so an Italian supplier's invoice and a Korean vendor's invoice both extract cleanly without switching any setting.