NVIDIA launches multilingual Nemotron OCR v2

NVIDIA has introduced Nemotron OCR v2, an optical character recognition model that can read documents in English, Chinese, Japanese, Korean and Russian with a single system. It can process up to 34.7 pages per second on an A100 GPU.

Optical character recognition, or OCR, converts an image containing text into content you can search, copy or analyze. It is what lets you digitize a scanned invoice, extract data from a form or read a table inside a PDF.

The problem was in the data

The previous version, Nemotron OCR v1, worked well in English but struggled with other writing systems. Its vocabulary included only 855 characters, which was not enough to properly cover Chinese, Japanese, Korean and Cyrillic scripts.

Expanding the set to 14,244 characters did not solve the problem on its own. The model could produce those symbols, but it had not seen enough examples to learn how they appear in real documents. The bottleneck was the lack of training data.

Creating millions of real images and labeling the words, lines, paragraphs and reading order in each one would be too expensive. NVIDIA chose to generate them instead.

Synthetic documents, exact labels

The team generated 12.2 million synthetic images: software-created pages with real text, different typefaces, backgrounds, colors, tables and multi-column layouts.

The advantage is that the system knows exactly what it placed in each image. That means it can automatically generate:

• The coordinates of each word, line and paragraph.
• The exact text appearing in each region.
• The relationships between words, lines and paragraphs.
• The correct order in which the document should be read.

To keep the images from looking too artificial, the process applied blur, noise, shadows, contrast changes, distortions and hundreds of font and layout combinations. The text came from mOSCAR, a multilingual corpus with content from 163 language subsets.

The result includes everything from text pages and presentations to tables, tables of contents, vertical columns and scenes with scattered words. In Chinese and Japanese, the model recognizes complete lines instead of words because those languages typically do not separate words with spaces.

The results

On the SynthDoG synthetic benchmark, normalized edit distance, a metric where a lower value means fewer errors, fell to these levels:

• Japanese: 0.046.
• Korean: 0.047.
• Russian: 0.043.
• Simplified Chinese: 0.035.
• Traditional Chinese: 0.065.

Nemotron OCR v1 recorded values between 0.564 and 0.923 in those languages, meaning the extracted text barely resembled the original.

On OmniDocBench, a test using real documents in English, Chinese and mixed-language content, multilingual Nemotron OCR v2 reached 34.7 pages per second. PaddleOCR v5 reached 1.2 pages per second in the same comparison, although the accuracy results are not directly identical across all scenarios.

The speed comes from its architecture. A single component analyzes the image, and its features are reused to detect and recognize text and understand the document's order and structure. This avoids repeating the most expensive part of the computation.

What changes for you

The model can process large volumes of invoices, contracts, forms and multilingual documents without first selecting a specific model for each language. It can also handle pages with tables, multiple columns or vertical text.

There is an important condition: the most notable results come from synthetic documents, and performance on real documents may vary depending on image quality, typography and layout. NVIDIA is publishing the model and dataset, so anyone can try them in a browser or integrate them into their own workflows.

The relevant advance is not just that the model knows more languages. It shows a practical way to add new writing systems using text and fonts, without manually labeling millions of images. The next point to watch is how much of that advantage it retains when faced with real, blurry documents or layouts that differ significantly from those used during training.