Low-complexity Real-time Single-channel Speech Enhancement Based on Skip- GRUs

Recently, algorithms based on deep neural networks have led to a significant speech enhancement performance improvement in terms of speech quality and intelligibility both for offline as well as online processing. However, obtaining a low-complexity and resource-efficient system is still a challenge. In this paper, we focus on real-time single-channel speech enhancement systems that are both compact and resource-efficient during inference. We propose two systems, either applying a real-valued or a complex-valued mask. Both systems are based on the Skip-GRU architecture, which employs a skip connection between the GRU layers. Experimental results on reverberant noisy signals demonstrate significant advantages of using the Skip-GRU architecture vs. the GRU architecture and applying a complex-valued mask vs. a real-valued mask. Moreover, the proposed Skip-GRU system with complex-valued masking achieves a similar speech enhancement performance as the best-performing baseline system but with a significantly reduced number of parameters and computational complexity.