In farming region of Inner Mongolia, animal husbandry is evolving from the traditional style to the modern style, which means the largescale sheep breeding, intensive management and industrial development. However, the newly extensive stable breeding facilities are easily to make sheep suffer from respiratory disease. In the early stage, cough sound of sheep can be detected for early disease warning and health diagnosis. In this paper, taking Dorper sheep, which has been widely promoted in Inner Mongolia, for an example, cough sound signal of sheep was automatically collected and recognized by computer. Without increasing the dimension of sound signal feature parameters, an improved Mel frequency cepstrum coefficient (MFCC) was put forward. The experimental results demonstrated that the 14dimensional parameters combined with improved MFCC, shorttime energy and zero crossing rate were used in the hidden Markov model (HMM) cough sound recognition system, whose recognition rate, error recognition rate and total recognition rate reached 86.23%, 7.17% and 88.43% respectively. And the combination parameters can be reduced to nine dimensions using principal components analysis (PCA) method. Furthermore, the cough sound recognition system based on HMM was enhanced by a backpropagation (BP) neural network, and it’s recognition rate, error recognition rate and total recognition rate reached 92.54%, 5.37% and 95.04%， respectively. Therefore, the recognition results meet the requirement of the Dorper sheep cough sound recognition.