Statistical voice activity detection in kernel space

Abstract

This paper proposes a statistical voice activity detection method in a high-dimensional kernel feature space by a nonlinear mapping. A Gaussian density model is presented using kernel principal component analysis to represent the nonlinear characteristics of the speech signal. The proposed approach offers a decision rule based on a multiple observation likelihood ratio test in the kernel space.

Enhanced voice activity detection in kernel subspace domain

“…6 Let x represent the clean speech vector and n be the additive noise vector, which is assumed to be uncorrelated with the clean speech vector. The observed noisy speech vector y is then given by y ¼ x þ n where x, n, and y are the D-dimensional vectors that are made up of time-domain samples in the original data space R D .…”

“…6) is proposed. A linear transform matrix to represent the kernel subspace is obtained by simultaneous diagonalization of the noisy speech and noise covariance matrices using KPCA.…”

“…The linear transform domain for VAD includes discrete Fourier transform (DFT), 1,2 discrete cosine transform (DCT), 3 Karhunen-Lo eve transform (KLT), 3 and signal subspace 4 domains, while the kernel feature space 5 is one of the nonlinear transform domains for VAD. 6 Recently, the kernel-based VAD algorithm in a nonlinear transform domain has been proposed and shown good performance. 6 The kernel-based VAD used multivariate Gaussian distribution (GD) and approximated covariance based on kernel principal component analysis (KPCA) (Refs.…”

“…6 Recently, the kernel-based VAD algorithm in a nonlinear transform domain has been proposed and shown good performance. 6 The kernel-based VAD used multivariate Gaussian distribution (GD) and approximated covariance based on kernel principal component analysis (KPCA) (Refs. 5 and 7) to represent the nonlinear characteristics of the speech and noise signals.…”

“…Many statistical model-based VAD algorithms using a likelihood ratio test (LRT) as a decision criterion have been developed in several linear and nonlinear transform domains. [1][2][3][4][5][6] These methods require the statistical models of the noise and noisy speech signals to describe their distributions more accurately in each domain. The linear transform domain for VAD includes discrete Fourier transform (DFT), 1,2 discrete cosine transform (DCT), 3 Karhunen-Lo eve transform (KLT), 3 and signal subspace 4 domains, while the kernel feature space 5 is one of the nonlinear transform domains for VAD.…”