Research · NewsShorts-Bench · Stage 5: Name-bar OCR
Reading TV news name bars in 13 languages: OCR models vs Gemini
Gemini 4 Argon reads 99.6% of names exactly and Cloud Vision 97.3%. Gemma 4 26B is the best open model at 93.8%; document-OCR leaders trail Gemini by 17-33 points and lose most on Arabic script and Devanagari.
- Data
- 520 name bars
- Compute
- H100 + CPU
- Truth
- Exact
- Metric
- Name exact %
Names read exactly (%) · higher is better · dot = measured · bar = 95% CI
When a news short shows someone speaking, the caption has to say who they are, and the safest source is the broadcaster's own name bar, the "lower third". Read it wrong and you put words in the wrong person's mouth. The OCR models topping document benchmarks like OmniDocBench should handle a line of text on a coloured bar. On real broadcast frames in 13 languages, most of them don't: they trail a frontier model by 17–33 points, and collapse on Arabic script and Devanagari.
Read the full study on Every Local AI