Low-Latency Streaming Target Speech Extraction for Edge Devices Using RWKV Linear Attention

Real-time streaming target speech extraction (TSE) on edge devices is severely hindered by the 'Dual Walls' of edge computing: the Compute Wall driven by algorithmic complexity, and the Memory Wall caused by high-frequency parameter fetching. This paper proposes an ultra-lightweight, compute- and memory-efficient streaming TSE system by reorganizing the Receptance Weighted Key Value (RWKV) architecture. By leveraging a linear attention formulation, the sequence modeling complexity scales at a constant O(1) during inference, dismantling the Compute Wall. Concurrently, a deterministic Short-Time Fourier Transform (STFT) front-end expands the frame shift to 10 ms, reducing memory bus traffic by 80% to overcome the Memory Wall. Under extreme -10 dB signal-to-noise ratio (SNR) conditions, our 3.18M-parameter model utilizes a conservative filtering strategy to preserve critical target formants, securing a robust Source-to-Interference Ratio (SIR) of 27.24 dB. Hardware benchmarking confirms the system achieves a Real-Time Factor (RTF) of 0.31 on an unaccelerated single-core CPU, demonstrating exceptional viability for resource-constrained tactical communications.