Noise compensation in speech recognition is a technique in which noise contamination is dealt with in the recognition phase rather than by some pre-processing system. In a previous paper, [1], we developed three noise compensation techniques for use in hidden Markov model based speech recognition. To date, however, no comprehensive examination and comparison of their performance has been carried out. This paper reports on such an examination and on the related topic of the use of a noise tracking "silence" model to "recognise" periods when there is no speech.