Enhancing Arabic Speaker Recognition with ECAPA-TDNN

This paper presents a fine-tuned Emphasized Channel Attention, Propagation and Aggregation - Time Delay Neural Network (ECAPA-TDNN) model for Arabic speaker recognition, with a focus on enhancing performance in noisy environments. The model was trained on the Voice of Celebrities 1 (VoxCeleb1) and VoxCeleb2 corpora combined with Arabic data from the Qatar Computing Research Institute (QCRI) Aljazeera Speech Resource (QASR), and was evaluated on the VoxCeleb1 test protocol (Vox1-O), the Arab Celebrity (ArabCeleb) dataset, a held-out QASR test split, and an in-house Arabic dataset of authentic recordings. Through targeted fine-tuning and data augmentation techniques, the proposed approach reduces the Equal Error Rate (EER) on Arabic datasets and improves robustness to noise, while maintaining satisfactory performance on English datasets. These findings indicate that careful adaptation can support the development of more balanced multilingual speaker verification systems, particularly for underrepresented languages such as Arabic.