“Towards single-channel blind dereverberation of speech from a moving speaker”
J. R. Hopgood and C. Evers
An overview of some key issues involved in single-channel blind dereverberation is presented, beginning with a discussion on the nature of time-varying acoustic channels, the problem of long acoustic impulse responses and an introduction to model-based solutions. The paper then discusses an approach in which the acoustic source is modelled by a time-varying AR process, and the channel by a linear time-varying all-pole filter. In each case, the time-varying filter coefficients are modelled as a linear combination of basis functions. Bayesian inference is used to estimate these coefficients. Although the propsed model needs validating against real acoustic sources and channels, it is proposed as an initial step towards solving the single-channel blind dereverberation.