Fine-Tuning Whisper for Speech Translation: A Case Study on Translating Darija to Arabic

The task of translating Darija, a dialect of Arabic, into Modern Standard Arabic (MSA) presents significant challenges due to linguistic variations and the lack of representation in major models. Traditional speech recognition and translation systems, like Whisper, are generally trained on more standardized languages and often struggle with dialects like Darija. To address this issue, we fine-tuned the small version of Whisper on the Darija-C corpus for speech-to-text translation into MSA. Our approach involved adapting Whisper's translation capabilities by fine-tuning the model with 5000 training steps and evaluating its performance using the BLEUmetric. The fine-tuned model achieved a BLEU score of 0.65, reflecting a marked improvement over the baseline. This research demonstrates the effectiveness of fine-tuning models for dialectal translation tasks, offering insights into how underrepresented dialects like Darija can be better supported by advanced speech translation systems.