Bilingual Speech Recognition On the Edge Using Machine Learning

Remarkable advancements in speech recognition systems have led to significant growth in assistive technology for individuals with disabilities. This paper signifies a collaborative effort to develop an embedded speech recognition system equipped with Machine Learning (ML) capabilities. The low-cost standalone speech recognition system is designed to assist individuals with disabilities by enabling them to operate their household appliances effortlessly using simple voice commands. One distinctive feature of the system is its ability to understand both English and Arabic, specifically the Emirati dialect. The development process involves the acquisition of datasets in both Arabic and English and the training of various ML models which will then be integrated into a microcontroller. Substantial testing is conducted to select the best ML model, ensuring high system accuracy. The model with 3 convolutional layers which is trained for 30 epochs is determined to be the optimal choice among the models, all of which demonstrated excellent accuracy, precision, and recall scores, exceeding 97%.