In this paper, we propose a novel lightweight speech enhancement network, MixSENet (Mixed Structure Speech Enhancement Network), designed to address the challenges associated with multi-scale feature processing of speech signals and coping with complex noise environments. The network is based on the U-Net architecture and innovatively introduces a hybrid structural block, which consists of a multi-scale parallel large convolutional kernel module (MSPLCK) and an enhanced parallel attention module (EPAM). The MSPLCK achieves large sensory fields and multi-scale feature extraction through parallel dilated convolution, whereas the EPAM can simultaneously process global shared information and local time-frequency features effectively coping with uneven noise distributions. Extensive experiments demonstrate that MixSENet shows significant performance advantages on the Voice Bank + DEMAND dataset, with a parameter count of only 0.43M, significantly reducing training and deployment costs. The method has substantial practical value in real application scenarios, especially for resource-constrained mobile devices and embedded systems.

A dual-stage training frame-level overlapping speech detection model is proposed in this paper, with its core being based on RepVGG blocks structure, and RepVGG is built on simple one-dimensional Convolutional Neural Networks (CNNs). Both the learning of local features and temporal features of acoustic signals are accomplished by the RepVGG structure. The model categorizes speech into triple classes: non-speech frames, non-overlapping speech frames, and overlapping speech frames. Firstly, RepVGG blocks are employed in the construction of a Speech Activity Detection (SAD) model to detect non-speech and speech frames. Subsequently, a Overlapping Speech Detection (OSD) model is constructed using RepVGG blocks and is trained with real labels to differentiate between non-overlapping and overlapping speech frames. Finally, these two models are combined and fine-tuned to directly output classification results. The proposed overlapping speech detection model achieves advanced performance on the AliMeeting dataset with fewer parameters.