A particle filter compensation approach to robust LVCSR

We extend our previous work on particle filter compensation (PFC) to large vocabulary continuous speech recognition (LVCSR) and conduct the experiments on Aurora-4 database. Obtaining an accurately aligned state and mixture sequence of hidden Markov models (HMMs) that describe the underlying clean speech features being estimated in noise is a challenging task for sub-word based LVCSR because the total number of triphone models involved can be very large. In this paper, we show that by using separate sets of HMMs for recognition and compensation, we can simplify the models used for PFC to a great extent and thus facilitate the estimation of the side information offered in the state and mixture sequences. When the missing side information for PFC is available, a large word error reduction of 28.46% from multi-condition training is observed. In the actual scenarios, an error reduction of only 5.3% is obtained. We are anticipating improved results that will narrow the gap between the system today and what's achievable if the side information could be exactly specified.