Detection of Weak Transient Signals Using a Broadband Subspace Approach

Abstract

We investigate the detection of broadband weak transient signals by monitoring a projection of the measurement data onto the noise-only subspace derived from the stationary sources. This projection utilises a broadband subspace decomposition of the data's space-time covariance matrix. The energy in this projected 'syndrome' vector is more discriminative towards the presence or absence of a transient signal than the original data, and can be enhanced by temporal averaging. We investigate the statistics, and indicate in simulations how discrimination can be traded off with the time to reach a decision, as well as with the sample size over which the space-time covariance is estimated.

A Polynomial Subspace Projection Approach for the Detection of Weak Voice Activity

“…In [14], a syndrome vector is computed by filtering the microphone signals through the eigenvector U ⊥ (z) ¡U ⊥ (n).…”

“…Instead of generating a syndrome vector in [14], a polynomial subspace projection P(z) = U ⊥ (z)U P ⊥ (z) ∈ C Q×Q is performed on the microphone signals x(n) to project them onto a reduced (Q − L) dimensional subspace. This will generate time signals y(n) with a reduction in energy contributions of the estimated L interferer components using…”

“…In [14], a broadband subspace-based approach is used to detect weak transient signals. The approach applies a polynomial matrix eigenvalue decomposition (PEVD), which is iteratively approximated by algorithms such as the secondorder sequential best rotation (SBR2) [15], [16] and sequential matrix diagonalization (SMD) [17], [18] in the time-domain or [19], [20] in the discrete Fourier transform (DFT)-domain, to generate the eigenvectors and eigenvalues.…”

“…The approach applies a polynomial matrix eigenvalue decomposition (PEVD), which is iteratively approximated by algorithms such as the secondorder sequential best rotation (SBR2) [15], [16] and sequential matrix diagonalization (SMD) [17], [18] in the time-domain or [19], [20] in the discrete Fourier transform (DFT)-domain, to generate the eigenvectors and eigenvalues. Filtering the signal through the eigenvector filterbank yields a syndrome vector, which is more discriminative towards detecting a transient signal [14].…”

“…In this work, we adapt [14] and investigate the idea of weak transient signal detection for VAD. The novel contributions of this paper are (i) a subspace-projection approach for VAD instead of the syndrome vector approach used for weak transient signal detection in [14], (ii) the use of realistic speech signals and measured room impulse responses (RIRs) and (iii) a comparison of the proposed approach against benchmark VAD algorithms in adverse environments. We first describe the goal of a VAD algorithm and provide a review of PEVD in Section II.…”