Robust speech/non-speech detection in adverse conditions using an entropy based estimator

This paper describes an original method for speech/non-speech detection in adverse conditions. We describe first how the theoretical dimension based on entropy can be used as a measure of the organisation degree of a signal and therefore adapted to speech/non-speech detection. The construction of an entropy based estimator called the local entropic criterion (LEC) is described and then tested on a speaker independent isolated digit database in white noise conditions. We show how to use the normalised theoretical dimension (NTD) and the LECs estimator, in order to perform a robust detection in adverse conditions. The results are comparable to those obtained in clean conditions at a signal-to-noise ratio (SNR) of 10 dB. We also show how it allows one to localise quasi-stationary parts of speech signals in conditions at an SNR of -25 dB