Noise-robust speech recognition using a cepstral noise reduction neural network architecture

The problem of speech recognition in the presence of interfering nonstationary noise is addressed. A method for noise reduction in the cepstral domain based on a universal approximator is proposed and tested on a large database of isolated words contaminated with nonstationary F-16 jet cockpit noise. The speech recognition system consists of a concatenation of an auditory preprocessing module, the cepstral noise reduction network (CNR network), and a neural network classifier. The proposed architecture performs a nonlinear autoassociative mapping in the cepstral domain between a set of noisy cepstral coefficients from the preprocessing module and a set of noise-free cepstral coefficients. The output from the CNR network is input to the neural network classifier, in which the output functions are approximations to the Bayes optimal discriminant functions. Noise reduction is possible in the preprocessing module and in the classifier, essentially making the system a three-stage noise reduction system. The average recognition rate on a test database was improved up to 65% when the CNR network was added to the speech recognition system