Abstract: Purpose Traditional organism recognition is based on the inherent contour, fingerprint, blood sample or unique biological characteristics of the organism. The implementation process is cumbersome and has obvious limitations. In order to establish a convenient and fast yak voiceprint recognition technology, this study used speech recognition technology to train the deep learning model of yak audio features. An improved voiceprint recognition yak organism feature algorithm based on ECAPA-TDNN network model is proposed. Method In this study, Datong yak and Qilian yak were taken as the research objects, the spectrum, Mel spectrum, LogMel spectrum, Mel cepstrum and LogMel features of different yaks were studied, and their audio features were trained by deep learning model. A voiceprint recognition yak organism feature algorithm based on ECAPA-TDNN network model was proposed. The BatchNorm layer and Dropout layer were introduced to improve the ECAPA-TDNN network to solve the noise and over-fitting problems. The ECAPA-TDNN network is improved, and the BatchNorm layer and Dropout layer were introduced to solve the problem of noise and over-fitting. At the same time, five different network models Dropout-capa-tdnn, Ecapa-tdnn, Panns-cnn6, Panns-cnn10, Panns-cnn14 and the same audio features (LogMel spectrum) were used for yak voiceprint recognition experiments. Result The results showed that the accuracy of yak individual identification is 96.11%, the accuracy of gender identification is 98.83%, and the accuracy of age group identification is 92.35%. Compared with the traditional model, it has higher accuracy. Conclusion This study provides an ideal model for individual identification of yaks and promotes the intelligent speech recognition technology of intelligent pastures, so as to provide technical support for promoting the high-quality development of intelligent animal husbandry in China.