Fact · Speech recognition
2percentage points
Kacprzak and Fraś found that the lack of number normalization can change WER by more than 2 percentage points on some datasets.
Checked against its source on October 17, 2026
In the source's own words
“may be substantial - more than 2 percentage points”
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, Abstract.
Open the original- What it measures
- The experiments compared normalization approaches on Polish parliamentary speech and on the VoxPopuli dataset.
- What it does not show
- A preprint on Polish and multilingual data. The authors say the gap was often larger than the differences between systems on popular multilingual benchmarks.