Compact Convolutional SNN Architecture for the Neuromorphic Speech Denoising

This paper introduces a novel approach to speech denoising within the framework of spiking neural networks (SNNs), specifically focusing on achieving a high signal-to-noise ratio (SNR) while minimizing power consumption and network size. The proposed SNN architecture leverages spatio-temporal dynamics by incorporating 3D Convolutional layers, a departure from the conventional 1D layers used in Temporal Convolutional Neural Networks (TCNNs). Additionally, the architecture inte-grates a Short-Time Fourier Transform (STFT) encoder and an Inverse STFT (ISTFT) decoder in conjunction with Sigma-Delta Spiking Neurons. This fusion of elements results in a compact SNN model that achieves commendable SNR performance while maintaining efficiency in terms of energy consumption and network size.