Peer-reviewed research
ASR Models for Traditional Emirati Arabic: Challenges, Adaptations, and Performance Evaluation
Maha AlBlooki et al., ACL Anthology (ICNLSP 2025 workshop paper), 2025
Read the originalRead on October 17, 2026
- What it is
- A 2025 study that built a small corpus of traditional Emirati Arabic from heritage broadcasts and literary sources, then tested five pre-trained speech models on it, zero-shot and after fine-tuning. It reports word and character error rates for Wav2Vec 2.0, XLS-R, Whisper Small, Whisper Medium and MMS.
- How much weight it carries
- Peer-reviewed workshop paper from a university research group. The corpus is small and covers one dialect, so the error rates describe that dataset, not every Arabic dialect. Its own authors note the dataset is low-resource.
Facts taken from this source
- In the 2025 Emirati Arabic study, fine-tuning the five speech models on the Emirati corpus improved both their word error rate and their character error rate.
“fine-tuning improves both Word Error Rate (WER) and Character Error Rate (CER)” (Abstract)
- In a 2025 study of traditional Emirati Arabic, zero-shot Whisper Small had a word error rate of 93.06 percent and Whisper Medium had one of 86.10 percent.
“with Whisper Small reaching 93.06% WER and 81.02% CER, and Whisper Medium slightly better at 86.10% WER and 75.01% CER” (Section 6 (Results), zero-shot baseline)