Deep Convolutional Neural Network for Arabic Speech Recognition.

Deep neural networks (DNNs) have made remarkable achievements in acoustic modeling for speech recognition. In this paper, we compare the performance of two proposed models based on Convolutional neural network (CNN). In the first model, CNNis used for features extraction and Long Short-Term Memory (LSTM) is used for recognition. In the second model, CNN with deep architecture is mainly used to execute feature learning and recognition process. This work is focused on single word Arabic automatic speech recognition. We explore the optimal network structure and training strategy for the proposed models. All experiments are conducted using the Arabic Isolated Words Corpus (ASD) database. The results demonstrate the performance and advantages of using the deep CNN for both features extraction and recognition steps.