Although most scholars have realized the importance of 3 D realistic sound embedded in future human computer interaction, there are still many obstacles to overcome. After intensive investigation of the current research situations, an improved HRTF (head related transfer function) model on the basis of neural network method was introduced to generate 3 D realistic sound. The HRTFs data employed can be reset by network training to satisfy a variety of requirements. Furthermore, with the utilization of neural network's nonlinearity approximation ability,all the HRTFs data(including magnitude characteristics and phase characteristics)at any incident angles can be obtained based on discrete sample values.Experiment and verification demonstrate the effectiveness and correctiveness of this scheme.