Bi‐DecoNet: Compact Dual‐Decoder Network for Robust Speech Enhancement Under Industrial Noise

This work presents Bi‐DecoNet, a compact dual‐decoder network for monaural speech enhancement in industrial noise environments. The framework integrates masking and spectral mapping strategies through a lightweight time‐frequency adaptive transformer to effectively capture temporal‐spectral dependencies. An attentional fusion module dynamically combines complementary decoder outputs to improve speech quality. Comprehensive evaluations on VoiceBank+DEMAND and TIMIT demonstrate state‐of‐the‐art performance with only 0.54M parameters and 2.1G FLOPs, achieving an optimal balance between enhancement quality and computational efficiency. The proposed method provides an industrially viable solution for speech enhancement applications.