When audio denoising meets spiking neural network

Abstract

Audio denoising techniques are essential tools for enhancing audio quality. Spiking neural networks (SNNs) offer promising opportunities for audio denoising, as they leverage brain-inspired architectures and computational principles to efficiently process and analyze audio signals, enabling real-time denoising with improved accuracy and reduced computational overhead. This paper introduces Spiking-FullSubNet, a real-time audio denoising model based on SNN. Our proposed model incorporates a novel gated spiking neuron model (GSN) to effectively capture multi-scale temporal information, which is crucial for achieving high-fidelity audio denoising. Furthermore, we propose the integration of GSNs within an optimized FullSubNet neural architecture, enabling efficient processing of full-band and sub-band frequencies while significantly reducing computational overhead. Alongside the architectural advancements, we incorporate a metric discriminator-based loss function that selectively enhances the desired performance metrics without compromising others. Empirical evaluations show the superior performance of Spiking-FullSubNet, ranking it as the winner of Track 1 (Algorithmic) of the Intel Neuromorphic Deep Noise Suppression Challenge.