Potential of Kolmorgorov Arnold Networks in Speech Enhancement Systems

ABSTRACT In order to capture complex, multiscale patterns, high-fidelity speech enhancement frequently calls for complex modeling. Despite adding nonlinearity, standard activation functions are not flexible enough to handle this complexity completely. An developing approach that uses learnable activation functions on graph edges, Kolmogorov-Arnold Networks (KAN), offers a possible substitute. In order to improve speech, this study examines two new KAN variations based on radial and rational basis functions. The radial variant is tailored to the 2D CNN-based decoders of MP-SENet, while the rational variant is integrated into the 1D CNN blocks of Demucs and the GRU-Transformer blocks of MP-SENet. The potential of KAN to enhance speech enhancement models is demonstrated by experiments conducted on the VoiceBank DEMAND dataset, which demonstrate that substituting KAN-based activations for standard activations enhances speech quality in both the time-domain and time-frequency domain approaches with negligible effects on model size and FLOP. Key Terms: Speech Enhancement, Kolmogorov-Arnold Networks ,Deep Neural Networks, MetricGAN .