Reproduce “PolyNorm: Few-Shot LLM-Based Text Normalization for Text-to-Speech” on OpenResearch

About this paper

Michel Wong, Ali Alshehri, Sophia Kao, Haotian He

Apple researchers developed PolyNorm, an LLM-based few-shot text normalization framework, which achieved significant reductions in Word Error Rate across eight diverse languages, outperforming a production-grade rule-based system. For instance, it reduced German WER from 10.74% to 4.17% and Italian WER from 15.02% to 4.56%, facilitating more scalable and efficient text-to-speech systems.

Key insights

• Large Language Models (LLMs) using few-shot prompting exhibit superior generalization capabilities for text normalization, consistently outperforming complex rule-based systems.
• LLMs effectively leverage their contextual understanding to accurately normalize ambiguous or dynamic inputs (e.g., URLs, symbols in different contexts) that challenge traditional methods.
• The LLM-driven framework significantly reduces the engineering effort and development overhead, enabling more cost-effective and scalable text normalization across many languages.

Summary by alphaXiv.