PolyNorm

This dataset accompanies the research paper, PolyNorm: Few-Shot LLM-Based Text Normalization for Text-to-Speech.

PolyNorm-Bench: Description

Text Normalization (TN) is a key preprocessing step in Text-to-Speech (TTS) systems, converting written forms into their canonical spoken equivalents. Traditional TN systems can exhibit high accuracy, but involve substantial engineering effort, are difficult to scale, and pose challenges to language coverage, particularly in low-resource settings. We propose PolyNorm, a prompt-based approach to TN using Large Language Models (LLMs), aiming to reduce the reliance on manually crafted rules and enable broader linguistic applicability with minimal human intervention. Additionally, we present a language-agnostic pipeline for automatic data curation and evaluation, designed to facilitate scalable experimentation across diverse languages. Experiments across eight languages showistent reductions in the word error rate (WER) compared to a produproduction-grade-based system. To support further research, we release PolyNorm-Benchmark, a multilingual data set covering a diverse range of text normalization phenomena.

Documentation

PolyNorm-Bench is a dataset with normalized and unnormalized data generated by DeepSeek-R1 as foundation across 8 locales, spanning 27 categories. Each synthetic example is edited and verified by internal language experts to ensure the highest standards of linguistic precision, naturalness, and overall quality. The orthography and formatting of synthetic data follow the conventions of target languages. There are 20 examples per category per language, resulting in 540 high-quality data points per language.

Normalization Categories

1. Cardinal
2. Date
3. Decimal
4. Ordinal
5. Fraction
6. Time
7. Currency
8. Unit (Measure)
9. Address
10. Acronym/Initialism
11. ISBN
12. Biological Classification
13. Roman Numeral
14. Telephone
15. Sports Score
16. Mathematical Expression
17. Symbol
18. Abbreviation
19. Chemical Formula
20. Legal Reference
21. Vehicle/Product Code
22. Geographic Coordinates
23. Version Number
24. License/Serial Number
25. Musical Notation
26. Stock Ticker
27. Electronic (URL/Email)

Languages

American EnglishGerman, French, Mexican Spanish, Italian, Lithuanian, Japanese, and Mandarin Chinese