Enhancing Accessibility of Feedback Collection in ML Models Through Keyword Spotting: A Moroccan Darija Model
Rim Miftah, Yassine Ibork, Asmaa Mourhir · 2024
This paper presents an innovative approach to enhancing feedback collection accessibility in machine learning models through the development of a keyword spotting system (KWS) designed for Moroccan Darija, a low-resource dialect. The solution allows blind and visually impaired users to give feedback on model outputs using voice commands, providing a hands-free method to indicate misclassification by saying “Mas7i7ch” (incorrect) or confirmation by saying “S7i7” (correct). Data collection involved gathering over 500 distinct audio samples in different environments, which formed the basis of the model training. The KWS model achieves 97.3% accuracy and 97% F1 score. We leveraged the KWS as part of the Human-AI Interaction of a machine learning application to collect user feedback about model predictions and misclassifications through voice commands. To evaluate our system's real-world applicability, a comprehensive user evaluation was conducted with over 112 participants of diverse genders, ages, vision conditions, and intellectual backgrounds. The system achieved a high accessibility rating of 88.4%, highlighting its effectiveness in improving inclusion and providing an intuitive voice-based interface for blind and visually impaired individuals.