Low detection accuracy of speech signal boundaries and pauses is one of the main problems of practical realization of diagnostic systems of psychogenic states. This paper proposes a noise-robust algorithm for 'speech/pause' segmentation, operating under free physical activity of a patient. In developing the algorithm the following methods were used: a method for adaptive processing of non-stationary signals – the Complementary Ensemble Empirical Mode Decomposition (CEEMD), a statistical data processing method – the Independent Component Analysis (ICA), a differentiation method using the concepts of normal distribution and one-dimensional Mahalanobis distance. The article presents a block diagram for the algorithm with a detailed mathematical description. The advantages over the known 'speech/pause' segmentation algorithms are shown. The developed algorithm enhances the actual detection rate by the average of 11.3%. A comparison of researches' results suggests that the developed 'speech/pause' segmentation algorithm is recommended for practical application in the diagnostic systems of psychogenic states, operating under free physical activity of a patient.