Dressing of grinding wheel is a process for tuning the conditions of the wheel working periphery. Therefore, sounds emitted during dressing may be closely related to wheel conditions necessary to obtain particular characteristics of machined surfaces. If the conditions of the wheel working periphery change only a little with grinding time, the grinding states will be predicted based on dressing sounds. The purpose of this paper is to construct a system for predicting surface finish and grinding forces based on dressing sounds. First dressing sound is processed by linear prediction method to obtain its feature of power spectrum as a vector with auto-regressive coefficients. Then a neural network is used to relate a feature vector to surface finish or grinding forces. As result, it is verified that a developed system is applicable to the prediction of the grinding states.