Japanese G2P (Parse Mode): leaderboard
Metric: Kana character error rate (%; can exceed 100) in parse mode (the LLM segments the sentence and gives each word's reading as JSON, then rule-based post-processing applies particle and long-vowel rules) on the 3,000 manually annotated sentences of the JVS nonpara30 corpus; kana character error rate against the reference readings after removing punctuation and normalizing long vowels; reasoning disabled (non-thinking mode, or low reasoning effort where it cannot be disabled), few-shot prompts; lower is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 37 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.6 | 0.52 |
| 2 | Gemini 3.1 Pro (Preview) (Low) | 0.62 |
| 3 | Claude Sonnet 4.6 | 0.82 |
| 4 | Gemini 2.5 Flash (Non-reasoning) | 1.08 |
| 5 | Kimi K2.5 (Non-reasoning) | 1.35 |
| 6 | GLM-5 (Non-reasoning) | 1.44 |
| 7 | GPT-5.2 (Non-reasoning) | 1.64 |
| 8 | Qwen 3.5 397B A17B (Non-reasoning) | 1.96 |
| 9 | GPT-OSS-120B (Low) | 3.03 |
| 10 | GLM-4.7 (Non-reasoning) | 3.7 |
| 11 | Gemma 3 27B | 5.75 |
| 12 | Qwen 3.5 35B A3B (Non-reasoning) | 6.12 |
| 13 | Qwen 3.5 27B (Non-reasoning) | 6.27 |
| 14 | Llama 3.3 70B Instruct | 6.58 |
| 15 | Qwen 3.5 122B A10B (Non-reasoning) | 7.57 |
Interactive version: theaggregate.ai/benchmark?slug=japanese-g2p-parse-mode · How It Works · Data refreshed daily, snapshot 2026-09-29.