Towards a practical and efficient WER Calculation for Arabic ASR systems
Mourad Abbas, Dhiya El Hak Megtouf · 2018
Word error rates (WER) achieved for Modern Standard Arabic speech recognition systems are much lower than those performed for English. Two essential factors are behind this fact. The first one is the intentional omission of diacritics in the Arabic scripts, and the other one is the agglutinative nature of Arabic which is not taken into account in the computation of WER. We focus on the second point by performing morphological analysis prior to claculalting WER, for both references and hypotheses. WER has been lowered by 4 % which can be considered as an encouraging result opening a new way for Arabic ASR evaluation.