The objective of this study is to investigate the feasibility of utilizing the signal features in vibration measurements during the milling process and the cutting parameters for predicting the surface roughness of S45C steel. The features of vibration signals are extracted by means of the envelope analysis, statistical computation, such as RMS (root-mean-square), kurtosis, skewness, and multi-scale entropy (MSE), as well as the frequency normalization. Through the correlation analysis, the features of higher priority are sifted out so that the prediction computation efforts can be reduced. The sifted vibration signal features are then collected as the input layer parameters of artificial neural network (ANN) for surface roughness prediction. The prediction results and accuracy through using different classes of input features are also discussed and compared. The experimental results show that the surface roughness is affected not only by the cutting parameters, but also by the vibration behavior during the milling process. Therefore, the cutting parameters combining the essential vibration features can be utilized to enhance the prediction accuracy of surface roughness during the milling process.