Accent-aware arabic speech processing using CNN–BiLSTM for dialect classification

Abstract

The Arabic language presents considerable dialectal diversity, which can pose challenges for speech technologies, particularly automatic speech recognition (ASR). Traditional models trained primarily on modern standard Arabic (MSA) do not account for regional accents, leading to poor performance on underrepresented dialects. However, this paper addresses this gap by introducing a hybrid CNN-BiLSTM model designed for automatic Arabic dialect classification. The dataset was divided into five categories gathered from various YouTube videos: Levantine (Jordan, Lebanon, Syria, Palestine), Gulf (Saudi Arabia, Yemen, Qatar, UAE), Nile Valley (Egypt, Sudan), Mesopotamian (Iraq, Kuwait), and North African (Algeria, Libya, Mauritania, Morocco). Therefore, a systematic augmentation was applied to address this imbalance, including pitch shifting, time stretching, and amplitude scaling, leading to 4,000 balanced utterances standardized to 16 kHz and 8 seconds. Convolutional neural network (CNN) layers captured localized spectral features, while bidirectional long short-term memory (BiLSTM) units modeled long-range temporal dependencies. Evaluation metrics, including classification report, confusion matrix, and learning curves, showed effective performance, with an overall accuracy of 88%.