One of the most required tasks in different applications of noisy speech processing are the measurement of signal-to-noise ratio (SNR) and precise simulation of noisy background. In this paper different methods of SNR estimation were compared with the reliability of these estimations from the noisy speech. It was found that segmental SNR seems to be the optimal solution for the measurement of the level of noise in the speech from the point of view reliability of the estimation procedure. The results with arithmetical averaging in segmental SNR evaluation are presented. This algorithm gives values close to standard global SNR and its big advantage is less sensitivity to the error of local SNR estimation.