This paper presents a comparative study of four steganalysis techniques for speech/audio files. The Mel-Frequency Cepstral Coefficients (MFCCs) are used for the acoustical analysis of the audio files. The following steganalyzers are assessed: Support Vector Machines (SVMs), Gaussian Mixture Models (GMMs), Deep Belief Networks (DBNs) and Recurrent Neural Networks (RNNs). These steganalysis methods are tested on three different steganographic approaches. Our experiments were carried out by using three steganographic techniques, namely, StegHide, Hide4PGP and FreqSteg that were applied to the Noizeus corpus. The results showed that the GMMs-based technique performed the best by reaching a perfect classification rate, without any error, followed by the SVMs, DBNs and RNNs.