Noise-aware Speech Separation Based on Dual-Path RNN Model

Deep learning-based speech separation has made remarkable progress recently, especially in time-domain methods such as dual-path recurrent neural networks (DPRNN). Despite these advancements, existing methods encounter significant challenges when addressing highly complex acoustic environments or meeting increasingly stringent separation quality requirements. In this paper, we propose improvements to the DPRNN framework, including optimized convolutional parallelism, an attention module for feature enhancement, and an improved residual module. Evaluations on the self-synthesized dataset and the widely adopted LibriMix benchmark show that our improved DPRNN outperforms the original model in terms of signal-to-distortion ratio (SDRi) and scale-invariant signal-to-noise ratio (SI-SNRi) with limited increase in model size. Its practical applicability is further verified in challenging environments, which will help develop efficient and robust speech separation systems.