Spike-Based Architectures for Energy-Efficient Audio Enhancement

Deep Neural Networks (DNNs) have shown remarkable success in speech denoising; however, their high computational and energy requirements make realtime deployment on edge devices challenging. In contrast, Spiking Neural Networks (SNNs) operate using sparse, event-driven spikes—offering a biologically inspired and energy-efficient alternative. In this study, we delve into a Spiking Neural Network (SNN) model that leverages the temporal dynamics of spiking neurons to capture long-range dependencies in the audio signal. By encoding the input audio into sparse and event-driven computations, the SNN can efficiently process the temporal information while requiring significantly fewer computations compared to DNNs. We present a real-time speech denoising system that maps noisy audio to sparse spike trains and processes them using diverse SNN architectures that aim to exploit time- and frequency-domain features. We investigated various baseline models and propose the Dual-Signal Transformation Spiking Network, a hybrid model that conducts frequency-domain enhancement through spectrogram masking and supplements it with raw waveform-based reconstruction in the time domain. Our experiments show that SNNs -- especially the Dual-Signal model -- are able to achieve competitive denoising performance while substantially lowering the computational cost—opening up possibilities for efficient and real-time auditory processing on neuromorphic hardware. This potentially contribute to the development of real time audio denoiser using SNN.