This paper proposes the online noise model adaptation technique using the modified quantile based noise estimation method for feature compensation of noisy speech that is based on the Gaussian mixture model for a robust speech recognition interface in real car environments. The proposed method is designed for an active online model adaptation method to cope with varying environmental noise conditions, and enhance speech recognition accuracy. This method is compensated on logarithmic filter-bank energies domain, and modified quantile based noise estimation method using beta-order harmonic mean is employed to the online noise estimation procedure. Experimental evaluation is done by using Aurora 2 speech database, and robust results were obtained than from other comparative algorithms.
Likelihood-Maximizing-Based Multiband Spectral Subtraction for Robust Speech Recognition
