Spoken American and Saudi Dialects Classification System Using Deep Learning Networks

Over the last few decades, deep learning systems have achieved significant success in speech-based interaction, including the classification of spoken dialects. Spoken dialect classification using deep learning systems aims to classify speech signals based on their features. In this article, we propose a spoken American and Saudi dialects classification system that compares Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BLSTM) classifiers with the Texas Instruments Massachusetts Institute of Technology (TIMIT) corpus and Saudi Accented Arabic Voice Bank (SAAVB) using a combination of carefully chosen speech features to uncover fruitful outcomes. After a series of experiments, the accuracies of the proposed systems using the CNN classifier for spoken American dialects classification with prosodic, acoustic, and Rhythm-Metrix features, and all their combinations, achieved 13.12%, 58.00%, 51.01%, and 68.50%, respectively. Likewise, we achieved classification accuracies of 9.06%, 32.03%, 28.78%, and 48.01% using the same features for spoken Saudi dialect classification. Compared with the proposed system using the BLSTM classifier, we achieved 21.83%, 56.03%, 65.63%, and 75.87% for spoken American dialects classification, and 10.18%, 13.89%, 47.20%, and 60.19% for spoken Saudi dialects classification, using the same features.