Preprint
The Hidden Cost of Digits: Number Normalization and WER in ASR Systems
Stanisław Kacprzak and Mieszko Fraś, arXiv (eess.AS), AGH University of Krakow, 2026
Read the originalRead on October 17, 2026
- What it is
- A 2026 preprint that measures how much WER changes when numbers are normalized before scoring, using Polish speech in VoxPopuli and a Polish parliamentary corpus.
- How much weight it carries
- A new preprint that has not yet been through full peer review. Its experiments are on Polish and multilingual data, so the size of the effect for English may be smaller, but the point that normalization choices move WER is shown directly.
Facts taken from this source
- Kacprzak and Fraś conclude that without number normalization, a comparison between two recognizers can be misleading.
“we have shown that the lack of number normalization can produce a misleading system” (Conclusion)
- Kacprzak and Fraś found that the lack of number normalization can change WER by more than 2 percentage points on some datasets.
“may be substantial - more than 2 percentage points” (Abstract)