|
Abstract:
Robustness to ambient noise and channel distortion
is crucial to real world speech recognition applications. I will
present a probabilistic graphical model of noisy speech generation and
discuss an approximate method for inference in this model, called Algonquin.
Algonquin uses an iterative variational method to find an approximation
to the posterior distribution of a clean speech vector, given the noisy
observed speech vector. Any model based cleaning method will perform poorly
if the parameters of the environment model (noise and channel) are poorly
estimated. The framework allows us to adapt the environment parameters
online using a Generalized EM method. I will present recognition results
for the basic and adaptive algorithms on the Aurora 2 database.