This paper addresses the problem of robust voice activity detection (VAD) capable for working at very low signal-to-noise ratios (SNR<10dB). A new algorithm that we propose is based on entropy estimation measures of the time-frequency magnitude spectrum. The problem of the estimation of the distribution of noise in detected non-speech segments of analysed signal is also presented. It is shown that the new entropy based VAD significantly outperforms the commonly used energy-based algorithms in all (stationary, non-stationary, white and coloured) noise conditions at SNRs from 10 dB down to -10 dB and below. One of the main advantages of the method proposed in this paper is that it is not very sensitive to the changing level of noise.