Spring Reverb Emulation with Hybrid Gated Convolutional Networks and State Space Models

Abstract

Modeling analog audio effects like spring reverbs is a long-standing challenge due to their complex, nonlinear behaviors, such as amplitude-dependent transients and long, dispersive reverberant tails. While deep learning has shown promising results, existing approaches often struggle to capture these characteristics simultaneously. In this paper, we propose GCN-SSM, a novel hybrid model architecture that combines a Gated Convolutional Network (GCN) with a State Space Model (SSM), leveraging sequential blocks of interleaved feedforward and recurrent layers. We evaluate the performance with spectral and time-domain losses, and a MUSHRA listening test. Our results show that both components are critical for achieving state-of-the-art perceptual quality. The GCN-SSM consistently outperforms the non-hybrid GCN across all metrics. With only 125.7k parameters and inference requiring 5.9 GFLOP for 1 second of audio at 44.1 kHz, the GCN-SSM is theoretically suitable for deployment on modern CPUs.