An Efficient Method of Flooring Threshold Calculation for Robust Speech Recognition Against Impulsive Noise

Performance of an automatic speech recognition (ASR) system tends to be dramatically degraded in the presence of impulsive noise. In the out pervious work (2003), we proposed flooring the observation probability (FOP) to compensate the adverse effect of impulsive noise on sensitive feature sub-vector. This paper describes a method that efficiently calculates the optimum flooring threshold, which is very crucial for the FOP algorithm. In this method the threshold is related to an approximate confidence region of the multivariate Gaussian distribution, thus the threshold can be calculated by Taylor series accumulation instead of complex multivariate integration. Tuning experiments show the calculated flooring threshold is very close to the optimum value, and evaluations in simulated impulsive noise also prove the efficiency of FOP algorithm with such calculated threshold