We have developed a vocal control method, based on forward and inverse models, to allow the anthropomorphic talking robot Waseda Talker No. 7 (WT-7) to produce various kinds of voices. The control parameters of the vocal cords on WT-7 are pressure, vocal cord tension and glottal opening, and the acoustic parameters are sound pressure, sound pitch and spectrum slope. The relationships among these parameters are complicated and difficult to model using conventional methods. Here we present a neural network (NN) control method. The learning process consists of creation of the NN forward model by back propagation methods and optimization of the inverse model using the forward model. In addition, a real-time auditory feed-back mechanism is used to reduce the error between the target and the generated acoustic parameters. Using this method, the control parameters can be adjusted to follow the target voice well.