Solutions for robust speech/non-speech detection in wireless environment

The use of speech recognition systems in noisy environments requires robustness to adverse conditions. An efficient detection of speech/non-speech segments is therefore necessary. Several approaches have been proposed in order to improve the robustness of speech/non-speech detection used for speech recognition in noisy conditions. In this paper, we describe a robust speech/non-speech detection algorithm based on the estimation of noise statistics: mean and variance. Results of several experiments carried out on a database collected over the GSM network show that this new approach improves the recognizer's global performances, especially in very noisy environments. Then, spectral subtraction is used as a preprocessing technique aiming to increase the robustness to noisy conditions. We show that the improvements concern mainly noisy conditions such as calls from outside or from running cars.