The segmentation of phonocardiogram (PCG) signals is the first step in the automatic diagnosis based on heart sounds. The majority of attempts to segment PCG signals depend on a reference provided by simultaneous electrocardiogram recordings. The algorithm proposed in this paper is based on the analysis of the PCG signal only and does not require an ECG reference signal. In this paper we propose the tracking of the log spectral components that vary slowly with frequency (the low-time components). That is Cepstral analysis is used to provide the features selected to represent the heart sounds. The algorithm utilises a hidden Markov Model to identify the S1 and S2 components of the heart sound, which delimit the systolic and diastolic cycles. The parameters of a simple hidden Markov model with single Gaussian distribution for continuous observations are learned from a training set of heart sounds. Once the parameters of the model are obtained PCG signals from different sets are used to test the segmentation procedure. (4 pages)