Comparing noise compensation methods for robust prediction of acoustic speech features from MFCC vectors in noise

The aim of this paper is to investigate the effect of applying noise compensation methods to acoustic speech feature prediction from MFCC vectors, as may be required in a distributed speech recognition (DSR) architecture. A brief review is made of maximum a posteriori (MAP) prediction of acoustic features from MFCC vectors using both global and phoneme-specific modeling of speech. The application of spectral subtraction and model adaptation to MAP acoustic feature prediction is then introduced. Experimental results are presented to compare the effect of noise compensation on acoustic feature prediction accuracy using both the global and phoneme-specific systems. Results across a range of signal-to-noise ratios show model adaptation to be better than spectral subtraction and able to restore performance close to that achieved in matched training and testing.