PolySpeech-100 - Low-Resource Languages: leaderboard
The 81 long-tail languages of PolySpeech-100 (Yang et al., KDD 2026): spoken Belebele reading-comprehension questions with four options, human-recorded where 2M-Belebele has them and neural TTS otherwise. Accuracy in percent (chance 25%), macro-averaged over the languages; higher is better.
Metric: Accuracy (%; spoken four-option reading comprehension from Belebele: the audio carries the passage, question and options; text instruction as system prompt; mean over 81 long-tail languages). Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 84.61 | #141 |
| 2 | Qwen2.5-Omni-7B | 40.18 | #1797 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=polyspeech-100-low-resource-languages · How It Works · Data refreshed daily, snapshot 2026-11-12.