Inversion-based nonlinear adaptation of noisy acoustic parameters for a neural/HMM speech recognizer