Kolloquium - Details zum Vortrag
Sie werden über Vorträge rechtzeitig per E-Mail informiert werden, wenn Sie den Newsletter des kommunikationstechnischen Kolloquiums abonnieren.
Alle Interessierten sind herzlich eingeladen, eine Anmeldung ist nicht erforderlich.
Kommunikationstechnisches Kolloquium: Bayesian Learning of Linear and Nonlinear Acoustic System Models in Hands-free Communication
Herr M.Sc. Sarmad Malik
14. September 2012
16:00 Uhr
Hörsaal 4G IKS
Linear and nonlinear acoustic echo control is an area of active research in acoustic signal processing. In this talk, DFT-domain system models will be discussed that are useful in describing the underlying linear and nonlinear physical processes.
It will be shown that the Bayesian learning of these models, which essentially amounts to a joint estimation task, can be carried out within a unified expectation-maximization framework. The focus will be on the optimization of the variational lower bound on the log-likelihood for obtaining equations of learning that are then executed iteratively to effectively cancel the acoustic echo. Due attention will be given to the involved mathematical operators, which are necessary for solving the optimization problem at hand.
Furthermore, a generalized derivation of a Bayesian post-filtering stage will be presented, which utilizes the parameters and distributions learnt in the echo canceller to suppress residual acoustic echo. The adopted approach yields closed-form and stable solutions that can incorporate all a prior beliefs and attain computationally efficient implementation by means of diagonalization.
