To solve the difficulty of low recognition accuracy caused by the difficulty of feature extraction in environmental sound signal recognition, a multi-scale parallel convolution network (MSPCN) model was proposed. Firstly, the two-dimensional convolutional neural network was used to deeply mine the time series features of the mel-frequency cepstral coefficients spectrogram in the sliding triangular window. It improves the information extraction ability of the model for sound features. Secondly, in order to capture the rich features in the sound signal more comprehensively and meticulously, this model utilized the residual networks with multi-scale inputs module, so that the model can focus on the information between feature channels. Thus, the ability to perceive the multi-level features of the sound signal from fine granularity to coarse granularity is realized. In addition, the regularization structure of dropout and residual is added in the forward calculation process of the model. These measures effectively enhance the generalization ability of the model and further improve the recognition performance. Furthermore, in order to verify the effectiveness of MSPCN, the Urbansound8K benchmark dataset was selected for testing. The results show that compared with the existing methods, the MSPCN model achieved a classification accuracy of 98.19% while maintaining a low computational cost, which shows the advancement and efficiency in the field of environmental sound recognition.