Steganography in audio files usually revolves around well-known concepts and algorithms, least significant bit algorithm to name one. This paper proposes a new, alternative approach where steganographic information is connected with the medium even more – by using the medium itself as the information. The goal of this paper is to present a new aspect of steganography, which utilizes machine learning. This form of steganography may produce statistically indeterminable steganographic files which are immune to brute force attempts at trying to retrieve the hidden messages. Then the proposed solution is verified against statistical analysis and brute force attacks with promising results.