Fourier magnitude-based steganography for hiding 2.4 kbpsMELP secret speech

In this paper, a novel steganographic technique based on hiding a secret speech, coded with a 2.4 kbps Mixed Excitation Linear Prediction (MELP) speech coder, in another (cover) speech bitstream, also coded with the MELP coder, is presented. First, the secret speech is embedded in the Fourier magnitude (FM) MELP parameter using an LSB method. Second, a necessary parameter for message extraction, the amount of embedded bits (AEB), is also transmitted in the stego speech using some of the fields of MELP frame. In addition, we provide a mechanism for re-synchronization in case of packet loss. The performance of the method is shown experimentally against a Support Vector Machine (steganalyzer). The accuracy of such steganalyzer is around 50%, i.e. equivalent to random guessing. Finally, we also provide transparency results for the stego speech. Although transparency is not usually a requirement in steganographic applications, the proposed method may also be used for other data hiding applications that require good fidelity.