Multi-Scale Channel Adaptive Time-Delay Neural Network and Balanced Fine-Tuning for Arabic Dialect Identification

The time-delay neural network (TDNN) can consider multiple frames of information simultaneously, making it particularly suitable for dialect identification. However, previous TDNN architectures have focused on only one aspect either the temporal or channel information, lacking a unified optimization both domains. We believe that extracting appropriate contextual and enhancing channels are critical Therefore, in this paper, we propose novel approach uses ECAPA-TDNN from speaker recognition domain as backbone introduce new multi-scale adaptive module (MSCA-Res2Block) to construct (MSCA-TDNN). MSCA-Res2Block is capable features, thus further enlarging receptive field convolutional operations. evaluated our proposed method ADI17 Arabic dataset employed balanced fine-tuning strategy address issue imbalanced datasets, well Z-Score normalization eliminate score distribution differences among different dialects. After experimental validation, system achieved an average cost performance (Cavg) 4.19% 94.28% accuracy rate. Compared ECAPA-TDNN, model showed 22% relative improvement Cavg. Furthermore, outperformed state-of-the-art single-network reported competition. In comparison best-performing multi-network hybrid competition, Cavg also exhibited advantage.