A scalable frequency domain-based linear convolution architecture for speech enhancement