Reading TV news name bars in 13 languages: OCR models vs Gemini

The hookLink to this section

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.

The setup, in one pictureLink to this section

60 public broadcast videos (VOA, White House, Al Jazeera) → background frames
   │ draw a name + role per locale (13 languages × 40)
   ▼
520 name bars with exact truth · 4 bar styles · 720p/360p · CRF 23/35
   ▼
9 readers: hosted APIs, open VLMs (1× H100), classic OCR (CPU)
   ▼
name exact % (95% CI) · role exact % · character error rate

The bars are synthetic, so the truth is exact: no labelling noise, no argument about what the bar says.

What the numbers sayLink to this section

• Gemini 4 Argon reads 99.6% of names exactly, Cloud Vision 97.3%. Every other model is behind Gemini by an interval that excludes zero.
• Gemma 4 26B is the best open model, at 93.8%, and the only open model that doesn't lose accuracy at 360p (93.7%).
• Classic OCR, told the language, beats most open VLMs. PP-OCRv6 with the language given reaches 85.8% on CPU, ahead of a 27B vision-language model. With its default model it manages 51%. The pipeline gets the language for free from speech recognition.
• Document-OCR leaders fall apart on non-Latin scripts. On Arabic script and Devanagari they are 41–99 points behind Gemini. GLM-OCR reads 1.2% of Arabic-script names.
• Resolution, not compression, separates open from closed. At 360p, OvisOCR3, Qwen3.8-27B and PaddleOCR-VL lose 16–20 points; the closed models lose at most 1.1. Heavy compression (CRF 35) costs everyone 6 points or less.

The dirty laundryLink to this section

• Fluent but wrong is the worst failure. On Arabic script GLM-OCR writes a different, plausible name instead of leaving the field blank. For attribution, a confident wrong name is worse than no name.
• Synthetic bars flatter everyone. Fonts are Noto and DejaVu; real bars animate, stack lines and carry logos. Earlier work found synthetic renders overstate Devanagari accuracy compared with real scans, so real frames could change the order.
• Gemini reads too much. It returns logos and background text too, and it is about 25 times slower than Cloud Vision, so it needs a step that picks out the name.

The recipeLink to this section

1. Hosted: Cloud Vision at 97.3% and 0.14 s per image, with Gemini 4 Argon as the accuracy ceiling.
2. Open: Gemma 4 26B, or PP-OCRv6 with the language passed in if you want CPU only.
3. Don't pick an OCR model from a document leaderboard for broadcast text. The ranking doesn't transfer.

What's nextLink to this section

A check on real broadcast frames, a step that separates the name from the role, and cost per broadcast hour.