A new approach for robust speech recognition using minimum variance distortionless response

In this paper, we proposed a new technique presents an extraction method for robust speech recognition using the MVDR (Minimum Variance Distortionless Response) spectrum of short time autocorrelation sequence which can reduce the effects of leftover of the nonstationary additive noise that remains after filtering the autocorrelation. To produce a further robust front-end, we present the customized robust distortionless constraint of the MVDR spectral estimation method through revised weighting of the subband power spectrum values based on the sub-band signal to noise ratios (SNRs), which adjust it to the new proposed technique. The new proposed functions allow the components of the input signal at the frequencies with minimum affected by noise to pass with better weights and attenuate more effectively the noisy and unwanted components. This revision results in decrease of the noise residuals of the projected spectrum from the filtered autocorrelation sequence, thus advancing to a more robust algorithm. Our proposed technique, when analyzed on Aurora 2 task for recognition applications, best performed all MFCC (Mel frequency cepstral coefficients) as the fundamental, respective autocorrelation sequence MFCC (RAS MFCC), and the proposed MVDR related features in numerous different noisy conditions.