A novel Adaptive Kolmogorov Arnold Sparse Masked Attention Model with multi-loss optimization for Acoustic Echo Cancellation in double-talk noisy scenario

Abstract

In recent years, deep learning techniques have emerged as the predominant approach for Acoustic Echo Cancellation (AEC), owing to their capacity to effectively model complex and nonlinear patterns. This paper presents a novel Adaptive Kolmogorov Arnold Network-Based Sparse Masked Attention Model (KASMA-LossNet) with multi-loss optimization inspired by the Kolmogorov Arnold representation theorem. The model is designed to capture complex nonlinear patterns, thereby improving speech quality and enhancing echo cancellation effectiveness, all while reducing the model’s computational load. The model effectively simplifies complex nonlinear multivariate functions into univariate representations, which is crucial for handling the intricate nonlinear aspects of echo. The KAN-based attention module is designed to apprehend dense speech patterns and analyze the relationships between echo, noise, and the target signal. It also excels at identifying long-range dependencies within the signal, assigning weight scores based on their relevance to the task, and offering exceptional flexibility, enabling the model to adapt to diverse acoustic conditions. To enhance training efficiency, three losses (smoothL1 loss, magnitude loss and log spectral distance (LSD) loss) are combined and integrated into the model, accelerating convergence, speeding up the training process, and delivering more precise results. The proposed model was implemented and tested, demonstrating notable improvements in echo return loss enhancement (ERLE) and perceptual evaluation of speech quality (PESQ). The reduction in computational load of the proposed system is demonstrated through steady GPU utilization and reduced convergence time.