Lightweight Streaming Speech Enhancement for AIoT-Enabled Wearable Hearing Aids Using Parallel Spiking Mamba

Abstract

Deploying deep learning-based speech enhancement in wearable hearing aids is challenging due to strict constraints on latency and real-time processing. Conventional models are often computationally intensive and unsuitable for resource-limited edge devices. To address this, we propose a lightweight and end-to-end streaming speech enhancement framework based on a parallel spiking Mamba architecture. The proposed model combines the computation-efficient, event-driven properties of spiking neural networks with the long-range sequence-modeling capabilities of state-space models. A parallel design enables simultaneous extraction of local acoustic features and global temporal dependencies, supporting efficient streaming inference. Experimental results show that the proposed method achieves 2.73 PESQ and 93.9% STOI in the streaming setting on the VoiceBank-DEMAND dataset, using only 0.26 M parameters and 0.11 G/s MACs. It also achieves an end-to-end algorithmic latency of 30 ms under the default configuration, demonstrating the potential of neuromorphic-inspired approaches for real-time speech processing in AIoT-enabled hearing aids.