This paper shows a research on the behaviour of the observation likelihoods generated by the central state of a silence HMM (Hidden Markov Model) trained for Automatic Speech Recognition (ASR) using cepstral mean and variance normalization (CMVN). We have seen that observation likelihood shows a stable behaviour under different recording conditions, and this characteristic can be used to discriminate between speech and silence frames. We present several experiments which prove that the mere use of a decision threshold produces robust results for very different recording channels and noise conditions. The results have also been compared with those obtained by two standard VAD systems, showing promising prospects. All in all, observation likelihood scores could be useful as the basis for the development of future VAD systems, with further research and analysis to refine the results.