Speech Enhancement via Metric GAN and Kolmogorov-Arnold Networks: A Deep Learning Approach in Python

Speech enhancement in noisy environments remains a critical challenge for robust voice communication systems. Traditional signal processing techniques and supervised deep learning models often struggle to generalize to diverse noise conditions and fail to optimize for human perceptual quality. This paper proposes a novel Metric GAN+KAN architecture, which integrates a Generative Adversarial Network (GAN) with Kolmogorov-Arnold Networks (KAN) to enhance speech signals by focusing both on perceptual fidelity and structural consistency. The GAN-based generator learns to map noisy speech spectrograms to clean counterparts, while the discriminator enforces perceptual realism. The KAN component introduces domain-aware constraints that preserve the harmonic structure and energy characteristics of speech. We train the system using perceptual metrics such as PESQ and STOI, enabling the model to directly optimize for intelligibility and clarity. Experimental results on the VoiceBank-DEMAND dataset demonstrate significant improvements over conventional methods, achieving a PESQ of 3.1, STOI of 0.88, and SDR of 15 dB. This work paves the way for real-time, intelligibility-focused speech enhancement systems in practical applications.