Speaker: Mathieu Hu
Date: October 5, 2017
Abstract:
The characterization of the room impulse responses via the cross-relation is reinterpreted for noisy conditions and exploited in this work to propose an approach for the blind identification of acoustic channels from reverberant noisy speech signals. In this novel approach, which aims to annihilate the speech content from the observed signal, a new objective cost function exploiting the non-stationary nature of speech against the stationarity of the additive noise is derived. Since room impulses are to be identified, the exponentially decaying tail and the spiky pattern of the early reflections are exploited to regularize the optimization surface. Through simulations, the proposed algorithm is shown to outperform the Robust NMCFLMS (RNMCFLMS) and the Variable Step-Size Unconstrained MCLMS in terms of accuracy of the acoustic impulse responses estimated from noisy reverberant speech signals for high Signal-to-Noise Ratios.
