MSFNet: A Nested Model for Multi-Sampling-Frequency Speech Enhancement

Recently, speech enhancement research has shifted towards universal models handling various input conditions under multiple degradations. While diverse training data improves robustness, handling data acquisition variations like sampling frequency (SF) requires architectural changes. This study focuses on SF dependency in networks involving fully-connected (FC) layers, as current fixed-resolution STFT-based methods are not directly compatible with FC layers. We propose a multi-SF network (MSFNet) incorporating an FC layer with the nested structure and learnable weight scaling (LWS) mechanism. LWS mitigates the distributional changes at the FC layer's output, and the nested structure allows compact submodel extraction. Our experiments demonstrate that the proposed method is computationally efficient and performs comparably to the independently trained models. Notably, an 8 kHz model has 33 times fewer parameters and 6 times fewer multiply-accumulate operations than a 48 kHz model.