Robust Voice Activity Detection Based on Noise Eigenspace

In this study, we propose a voice activity detector (VAD) based on a noise eigenspace, which improve the robustness of VAD by utilizing the compression capability of the eigenspace. A noise eigenspace is constructed by using eigenvalue decomposition of the noise correlation matrix. When noisy speech is projected into the noise eigenspace, the noise energy is packed into a few dimensions with large eigenvalues, and those dimensions hopefully possess relatively less speech, because the speech energy distribution is usually different from noise energy distribution. The noise can be reduced by discarding those dimensions with large noise energy, while no significant loss occurs in speech. To track noise variation, the noise eigenspace is periodically updated, where the computation cost for eigenspace construction can be kept at an acceptable level. The proposed VAD was evaluated using the TIMIT database mixed with several noises. The experiment showed that the proposed VAD is more accurate than previous VADs in noisy environments.