Digital steganography introduces statistical distortion to some extent. Thus, steganalysis can be used to classify an object with or without hidden information. In this paper, we present a novel approach to detect the presence of LSB (Least Significant Bit) steganographic messages in the voice secure communication system. A distance measure, which has been proved to be sensitive to LSB steganography by ANOVA (analysis of variance), is denoted to estimate the difference between the host signal and the stego signal. Then a ML (maximum likelihood) decision is combined to form the classifier. Statistical experiments show that the proposed approach has highly accurate rate and low computational complexity.