Abstract
We propose a novel method for broadband noise reduction in speech, based upon a combined multiway subspace data reconstruction with adaptive signal enhancement. The multiway subspace method exploits a higher-order representation of the observed channel-signals in time-domain only. In the noise reduction context, multiway subspace data reconstruction is firstly performed using observed noisy signals, in order to estimate the overall shape of the original speech source signal. The enhanced speech signals are then obtained via post-processing by adaptive signal enhancers. Simulation studies show that the proposed two-staged approach can yield better noise reduction in terms of both objective and subjective performance metrics, compared to a widely used nonlinear spectral subtraction method and the method using sliding subspace projection.
