Nurse grunting of lactating sows, as a signal to attract piglet sucking, the frequency of which has a fixed relationship with the motive of sows milk secretion, will benefit for building a feeding pattern between the sow and its’ piglets. As an important trait of good maternal behavior, nurse grunting can be one of the criterion in maternal behavior. However, the noises from environment and piglets screams often disturb the recognition and extraction of time and frequencydomain characteristics. The nurse grunting of Xiaomeishan sows combining with the noises from environmental interference noise and piglets’ screaming calls were used in this study. Firstly, according to the energy of noise is evendistributed in the frequency domain, which corresponds to the maximum entropy characteristics, the layers of wavelet packet was selected, and the entropy of corresponding wavelet packet coefficient was calculated. Secondly, the power ratio was calculated to distinguish the types of noise. Finally, the different denoising methods in sound signal with different kinds of noise were compared. The results showed that the main frequency range of a nurse grunting was ranged from 0Hz to 800Hz, while the frequency of piglets’ screaming calls was ranged from 1000Hz to 8000Hz, and the frequency of other interference noise was distributed from 0Hz to 8000Hz. The power ratios of 0Hz to 1000Hz in pure nurse grunting without noise, with environmental noise, and with screaming calls were 1, between 0.85 to 0.97, and between 0 to 0.8, respectively. Meanwhile, with db10 as a wavelet packet basis, the 5 layers and 3 layers wavelet packet transform were used, the main noise ranging from 0Hz to 8000Hz and residual noise ranging from 1000Hz to 8000Hz can be removed. Therefore, the reconstructed signal can be used to analyze the feature of nurse grunting in Xiaomeishan sows.