LLMs benchmarked for Japanese Grapheme-to-Phoneme conversion

A new study benchmarks over 30 large language models (LLMs) for Japanese grapheme-to-phoneme (G2P) conversion, a crucial step for text-to-speech systems. Researchers compared LLM performance against traditional morphological analyzers using two prompting strategies: a parse mode involving morphological analysis and rule-based conversion, and a direct mode where LLMs predict kana readings. The findings indicate that LLM size, version, and specialized Japanese training significantly impact results, with the top LLMs achieving a kana character error rate below 0.52%, outperforming the best conventional tool. The parse mode generally yielded better results due to rule-based post-processing, and using LLM-predicted kana with a Text-to-Speech system improved pronunciation. AI

IMPACT This research highlights the potential of LLMs to improve Grapheme-to-Phoneme conversion accuracy, which could lead to more natural and robust text-to-speech systems, particularly for languages with complex phonetic rules.

RANK_REASON Academic paper detailing benchmark results for LLMs on a specific NLP task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Japanese
• large language models
• Text To Speech

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →