Voice activity detection using discriminative restricted Boltzmann machines

Voice Activity Detection (VAD) plays an important role in current technological applications, such as wireless communications and speech recognition. In this paper, we address the VAD task through machine learning by using a discriminative restricted Boltzmann machine (DRBM). We extend the conventional DRBM to deal with continuous-valued data and employ feature vectors based either on mel-frequency cepstral coefficients or on filter-bank energies. The resulting detector slightly outperforms the VAD often used as benchmark for detector comparison. Results also indicate that DRBM is able to deal with strongly correlated feature vectors.