In this work, we explore the potentials of predictability as an objective measurable quantity to identify elements of speech which are most important for intelligibility. Specifically, we propose a measure which is a non-intrusive perceptually relevant novel estimate of the information theoretical quantity 'mutual information'. This measure is computed for discrete time frames of speech signals and utilizes deep convolutional neural networks. In a listening test, the proposed measure was compared to two existing methods, namely sound intensity and cochlea-scaled spectral entropy. We found that - in certain conditions - the proposed measure better identified time frames important for speech intelligibility compared to sound intensity and cochlea-scaled spectral entropy. However, in other conditions, the proposed measure failed to identify such frames. The results suggest potentials of predictability as an objective measure, however, alterations should be made to the proposed measure in order to more sensible evaluate the measure.