PolyNorm: Few-Shot LLM-Based Text Normalization for Text-to-Speech

Michel Wong, Ali Alshehri, Sophia Kao, Haotian HeEMNLP 2025

TL;DR

This paper introduces PolyNorm, a novel few-shot LLM-based approach to text normalization (TN) for Text-to-Speech (TTS) systems. By leveraging large language models with in-context learning, PolyNorm significantly reduces the need for manual engineering and offers a scalable, language-agnostic solution that outperforms traditional rule-based systems across multiple languages.

What Problem Does This Paper Address?

Text Normalization (TN) is a crucial preprocessing step in Text-to-Speech (TTS) systems, converting written text (like numbers, abbreviations, and special characters) into their spoken equivalents. Traditional TN systems, often built on complex rule-based approaches (e.g., WFSTs), are highly accurate but require substantial manual engineering effort, are difficult to scale to new languages, and struggle with low-resource languages or morphologically rich ones. Furthermore, the same written form can have different spoken interpretations depending on context and language, making a universal TN system challenging. This paper aims to address these limitations by developing a more efficient, scalable, and linguistically adaptable TN solution.

Methods

The core of PolyNorm is a few-shot prompting strategy using Large Language Models (LLMs). Instead of relying on extensive hand-crafted rules, PolyNorm employs in-context learning (ICL) examples to guide the LLM. The prompt design is structured and language-agnostic, consisting of three components:

  1. Instruction Prompt: Defines the TN task.
  2. In-Context Learning Examples: Demonstrates how various text types should be normalized, tailored to specific languages.
  3. Target Unnormalized Input: The text to be normalized.

This approach allows the LLM to generalize from a small number of examples, eliminating the need for extensive fine-tuning or rule-based engineering. The authors also developed a language-agnostic pipeline for automatic data curation and evaluation to facilitate scalable experimentation. They iteratively refined the system by analyzing errors and updating ICL examples.

Datasets

The research utilizes a subset of the English Kestrel dataset for initial exploration. However, to address the limitations of existing datasets and create a truly multilingual benchmark, the authors introduce PolyNorm-Benchmark. This new dataset comprises unnormalized and normalized text pairs across 27 normalization categories, built upon Kestrel categories and expanded with new ones. It includes 20 examples per category, totaling 540 high-quality data points per language. The data was generated using DeepSeek-R1 and then meticulously edited and verified by internal language experts to ensure linguistic precision and quality.

Key Insights and Ideas

  • LLMs as a Scalable TN Solution: The paper demonstrates that LLMs, when guided by few-shot prompting and ICL, can achieve high-quality text normalization without extensive manual rule creation, significantly reducing engineering effort.
  • Language-Agnostic Prompting: The core prompt structure is designed to be language-agnostic, with language-specific nuances handled through localized ICL examples, enabling broad applicability.
  • Context is King for Ambiguity: The paper highlights how LLMs excel at resolving contextual ambiguities in TN, a significant challenge for rule-based systems, as illustrated by the "17°" example.
  • Data Curation is Key: The creation of the PolyNorm-Benchmark dataset, with expert verification, is crucial for robust evaluation and highlights the importance of high-quality data for LLM performance.
  • Iterative Refinement: The error analysis and hillclimbing iteration process, where ICL examples are updated based on observed errors, is a powerful method for improving LLM performance on specific tasks.

Key Results

PolyNorm, using GPT-4o-mini and GPT-4o models, demonstrated significant improvements over a production-grade rule-based system across eight languages (German, American English, Mexican Spanish, French, Italian, Lithuanian, Mandarin Chinese, and Japanese).

  • Word Error Rate (WER): GPT-4o achieved an overall WER ranging from 4.17% to 7.88% across languages, outperforming the baseline system. For example, in German, GPT-4o achieved a WER of 4.17% compared to the baseline's 10.74%.
  • BLEU Scores: PolyNorm also showed improvements in BLEU scores, indicating better surface similarity to the expected spoken form.
  • Flexibility and Coverage: The LLM-based approach showed substantial advantages in flexibility, coverage, and handling previously intractable normalization challenges with minimal manual intervention.

Limitations and Future Work

Limitations: * Dependence on ICL Quality: PolyNorm's performance is highly dependent on the quality and representativeness of the data used for in-context learning. Categories not well-covered in the examples may lead to poorer performance. * Error Propagation: Errors in normalization can propagate to downstream tasks like translation or sentiment analysis. * Edge Cases: Despite expert-reviewed examples, LLMs may still struggle with uncommon linguistic patterns or edge cases.

Future Work: * Diacritization Restoration: Enabling diacritization restoration in languages like Arabic and Hebrew, where diacritics are crucial for meaning. * Integrated Tokenizers & Multitask Learning: Exploring integrated tokenizers and multitask learning for non-whitespace languages like Japanese. * Suprasegmental Features: Incorporating suprasegmental features like pitch accent and tone prediction into the normalization pipeline for more natural speech synthesis.

Summary updated

BibTeX
@inproceedings{wong2025polynorm,
  author = {Michel Wong and Ali Alshehri and Sophia Kao and Haotian He},
  title = {{PolyNorm: Few-Shot LLM-Based Text Normalization for Text-to-Speech}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure