In the process and management of breeding birds breeding, with the help of noncontact and continuous sound detection as well as some intelligent equipment, the breeder can fully understand the health status and individual needs of breeding birds, which can improve production efficiency as well as animal welfare. A kind of lightweight convolution neural networks for breeding birds voice recognition was proposed. The sound of the Hy-line brown breeding birds was taken as the research object, and five kinds of common sounds in the breeding bird house were collected, then the one-dimensional signal of sound was converted into two-dimensional image signal. Based on the great advantages of convolutional neural network in image recognition, a lightweight deep learning model was established, with 80% data as training and 20% data as testing. This model realized the efficient detection process of animal sound signal from input to output of recognition results. By comparing and analyzing the recognition methods of previous studies, the proposed method greatly improved the overall accuracy rate of recognition of five kinds of common sounds in breeding birds' house by 3.7 percentage points. The experimental results showed that the average accuracy rate of this method was as high as 95.7%. The recall rate of the model for drinking water, fan noise and laying call were all up to 100%, and the precision rate and F1 value of fan noise and laying call were also up to 100%. While, the recall rate of stress call was the lowest value of 88.3%. The research result provided some theoretical reference for the research and development of unmanned intelligent equipment in the future large-scale chicken house.