Search

The size of the receptive field of the convolutional neural network is related to the size of the convolution kernel. And the traditional convolution uses a fixed-size convolution kernel, which limits the feature perception ability of the network model. In addition, due to the parameter sharing mechanism of the convolutional neural network, it used the same feature extraction method for all pixels in the spatial region. However, there are differences in the distribution of noise signals and clean speech signals in the noisy spectrogram, especially in the complex noise environment, the general convolution method is difficult to achieve high-quality speech signal feature extraction and choosing. In order to solve the above problems, this paper proposed a multi-scale region adaptive convolution module, which used multi-scale information to improve the feature perception ability of the model and automatically allocated the area adaptive convolution achieve and improved the denoising ability of the model. The experiments on the TIMIT public datasets show that the proposed algorithm has achieved satisfactory results in the metrics of speech quality and intelligibility. [ABSTRACT FROM AUTHOR]

In order to improve the information processing ability of the model in complex scenes, this paper proposed a dynamic selection mechanism based on attention, which selectively allocated weights according to the current information, effectively fused the feature outputs of deformation convolution and ordinary convolution, and adaptively balanced deformable convolution and standard convolution, so as to improve its representation ability. In addition, the learning ability of the model was further enhanced by means of iteration without additional parameters. It used seven different kinds of noises from Noise-X92 in TIMIT corpus, and carried out experiments in various SNR environments. The results show that the proposed algorithm outperforms other recent deep learning algorithms in terms of intelligibility and speech quality, regardless of SNR and whether noise appears in the training data set. [ABSTRACT FROM AUTHOR]