Enhancing Automatic Speech Recognition for Moroccan Darija Using Transformer Models

Abstract

Automatic Speech Recognition (ASR) for the Moroccan Darija a dialect characterized by high linguistic variability and a lack of standardization is a challenging task. We fine-tuned two state-of-the-art transformer-based models, Wav2Vec2-XLSR-53 and Whisper, and created a 5-hour high-quality Moroccan Darija corpus extracted from public YouTube channels. The results show that, in noisy conditions, Whisper is more accurate; however, Wav2Vec2 outperforms Whisper in clean conditions, reaching a Word Error Rate (WER) of 10.64% and a Character Error Rate (CER) of 2.22%. This work can help determine how to set up an effective ASR system for Darija and other low-resource, heavily code-switching dialects.