Advancing Moroccan Darija Digit Recognition Through a Deep Learning Approach with RNN, LSTM, and GRU Models

In the last six decades, Automatic Speech Recognition (ASR) technology has revolutionized daily interactions between humans and machines, primarily through voice interfaces. ASR applications like virtual assistants (Cortana, Alexa, Siri, Google Assistant), voice dictation, and more, have significantly eased daily routines, supporting multiple international languages like French, English, Arabic, and Spanish. However, limited research has been dedicated to ASR in language dialects such as the Moroccan Darija. The Arabic language, with its numerous dialects, presents unique challenges, including variations in pronouncing the ten digits from zero to nine. In this article, we present a speech recognition (SR) model for Moroccan Arabic digits, based on Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). The designed SR System represents the input signal as a Mel-Frequency Cepstral Coefficients (MFCC) vector by feature extraction after processing the signal to reduce noise. The GRU cells can solve temporal dependencies problems that require long-term learning and vanishing problems related to RNN. The constructed dataset for Moroccan Darija digits contains 2661 samples collected using a form to record the different digits separately. We used 80% of the samples for training, 10% for training validation, and the remaining 10% for testing. The GRU model achieved a training accuracy of approximately 93% and a testing accuracy of 74%.