Complexity-Driven and Path-Selective Decoding for DPRNN-based Signal Separation

In complex communication environments, the deep learning-based signal separation method Dual-path RNN(DPRNN) encounters significant challenges due to the inherent difficulty of separating complex mixed signals and the pronounced performance bottlenecks in its decoder architecture. To overcome these limitations, this paper introduces a complexity-driven and path-selective decoding framework, incorporating an adaptive decoder that dynamically adjusts to signal complexity. The proposed decoder employs a complexity estimator to dynamically assess signal complexity in real time, enabling adaptive path selection. Furthermore, it integrates a feature reconstruction and balancing module to establish a feature enhancement path, thereby improving the representation of complex signals. Experimental results demonstrate that under ideal channel conditions, the proposed decoder reduces the error rate (ER) by an average of 63.9% compared to the baseline DPRNN model across mixed-signal separation tasks, while preserving signal fidelity. Ablation studies validate the effectiveness of the complexity-driven mechanism and feature enhancement path, underscoring the necessity of the complexity estimator for improving model generalization. This research offers novel insights and methodologies for designing efficient, high-performance signal decoder tailored to complex communication environments.