We collected videos from YouTube featuring Egyptian Arabic speakers from various channels and diverse content. These videos were selected to ensure a wide range of speech patterns, accents, and topics. Importantly, each video was accompanied by subtitles or closed captions.

Audio Extraction

We extracted the audio track from each video using appropriate tools. This allowed us to isolate the spoken content while disregarding the visual aspects. Subtitles as Labels

The subtitles or closed captions provided with the videos were utilized as the ground truth labels. These subtitles accurately represented the speech content in written form. By separating the audio from the video and utilizing the subtitles as labels, we were able to create a dataset suitable for training and evaluating our speech-to-text model. This approach facilitated the inclusion of diverse spoken content, ensuring that our model is capable of accurately transcribing Egyptian Arabic speech across various topics and contexts.

The collected dataset serves as a valuable resource for training and evaluating our models, enabling them to achieve high accuracy and robustness in converting spoken Arabic language into written text.