When people communicate with telephone type systems, it is often assumed that the listener would notice any modification of the speaker's voice. It is possible however to introduce small changes that would not be noticed by a listener but could modify the reading of a Voice Stress Analyser, popularly referred to as a lie detector. Existing approaches to checking the integrity of voice require significant amounts of processing or are able to detect only non-subtle modification such as change of speaker. With the advent of real time voice modification using software and hardware based signal processing, we argue that it is not possible at present to easily determine if a received unencrypted voice message has been modified in some subtle way. This is particularly important in the current climate of biometric and psychophysiological analysis. We critically investigate alternative approaches to a voice integrity check in the light of the subtle changes that might be made to a speaker's voice and propose a method that enables voice integrity checking for digital communications in a variety of scenarios.