Abstract: As a standard of speech compression coding for mobile Internet，AMＲ has been widely used，and also provides a new carrier for steganography． Due to the inaccurate prediction of pitch delay parameters，the existing steganographic algorithms hide the information by adjusting the pitch delay parameters． In this paper，we find the correlation between the pitch delay of each subframe by analyzing the characteristics of the pitch delay prediction coding of AMＲ encoding algorithm． Thus，we propose a new steganalysis algorithm which consists of features based on first-order Markov transition probability in subframe group and the second-order differential Markov transition probability feature of pitch delay． The experimental results show that， in the mixed training environment， the accuracy of the proposed algorithm is significantly improved，compared with the existing algorithms，especially in the case of low embedding rate， the performance is improved significantly． In the case of 10% relative embedding rate， the detection accuracy of stego samples is improved by 1% ～10% compared with the existing algorithms． 

Key words: AMR, pitch delay, Markov transition probability, steganalysis