Efficient Speech Denoising via Frequency-Domain Reservoir Computing

Speech denoising for real-time, edge-device applications remains challenging due to the high computational and training demands of conventional deep learning models. In this work, we propose a hardware-efficient denoising system based on a frequency-domain reservoir computing (RC) framework; cornerstone framework for Neuromorphic Computing. Unlike traditional deep architectures requiring extensive training, our approach uses a fixed, randomly connected 500-neuron reservoir with training limited to a linear readout, significantly reducing computational demands and facilitating efficient hardware deployment. The denoised signal is then reconstructed by combining the predicted clean magnitude with the original noisy phase via inverse STFT. Experiments on the LibriSpeech dataset yield an SNR of 6.31 dB, a PESQ of 2.74, and a STOI of 87.8%, demonstrating the effectiveness of our method for resource-constrained, real-time speech enhancement.