I. Introduction
The process of classifying conversational speech into speech and non-speech periods is called voice activity detection (VAD). Such algorithms are important in many speech-processing applications such as speech recognition [1], speech enhancement [2], wireless communication [3] and speech coding [4]. The basic VAD algorithm consists in extracting features from an input signal and through the application of a threshold, classify speech segments as either speech or silence periods. Such an algorithm operates well in noise free environments, however in real life applications, such as in wireless devices, speech signals are distorted by background noise. In mobile applications, the signal-to-noise ratio (SNR) can vary from + 30dB to −10dB and the background noise types can vary from stationary to nonstationary abruptly [5]. In most applications, incorrect classification of speech periods can cause a serious deterioration of the speech quality, consequently noise robust VAD algorithms that operate reliably at very low SNR need to be developed.