← Data and charts
AI & evaluation · Updated 2026-11-04

Removing combining marks cuts Yoruba word error from 64.7% to 33.5%, but changes what is being measured

Word error rate (WER) of the language-specific speech models, with and without combining marks. Lower is better.

Hover, focus or click a row. Use the legend to show or hide a series.

Unit: % WER. 100 clips per language. Removing combining marks also removes underdots, not only tone marks, so the two conditions measure different notions of equivalence.

Cite this chart

Extrafemi Research (2026). Removing combining marks cuts Yoruba word error from 64.7% to 33.5%, but changes what is being measured [Chart]. Data from Naija Eval: An Open Evaluation Pipeline for Nigerian-Language Translation and Speech Recognition. https://research.extrafemi.com/data/naija-eval-speech-error-rates