A Lightweight Dual-Path Conformer Network for Speech Separation

In recent years, the combination of CNNs and transformers has been widely used in speech processing tasks, resulting in a significant improvement in the modeling performance of models for global features. However, with the stacking of convolutional blocks and the size of the convolutional kernel, the model parameters continue to increase, resulting in increasingly serious consumption of computer resources. Therefore, in this paper, a multiscale dual path Conformer network based on depth-separable kernel convolution is proposed, taking into account the network model separation performance while minimizing the consumption of computer resources by model parameters. The main innovations are the use of adaptive multiscale depth-separable convolution in the encoder part so that the model can adaptively adjust the convolution kernel size according to the input features to obtain multiscale features and the use of a depth-separable kernel convolution module in the separation network part so that the model can make small changes in the model parameters while increasing the sensory field by using a large convolution kernel. The experimental results show that the parameters and performance of the proposed network model achieve better results in both pure speech separation and speech separation tasks with noise and reverberation.