Arabic Speech Commands Recognition with LSTM & GRU Models Using CUDA Toolkit Implementation

Abstract

Speech commands recognition, especially in the Arabic language, is a fertile field for research, due to the severe shortage of datasets in the Arabic language, unlike the English language, available on the most famous set of data for spoken commands, which is the Google Speech Commands dataset. This has sped up research and given rise to numerous fresh deep-learning methods involving keyword discovery. This paper introduces the classification of two classes of Arabic speech commands taken from the Arabic Speech Commands Dataset (v1.0) using LSTM & GRU models. Furthermore, the model will be trained on a GPU using NVIDIA's CUDA to reduce training time. Several experiments were run during training to investigate how various factors affect the system's performance and ensure that our model's parameters are the best. The outcomes reveal that the proposed method had good training, validation, and testing accuracy, and because of the innovation (CUDA Toolkit), the training was completed quite quickly, and the model accurately identified the commands.