New Features Using Robust MVDR Spectrum of Filtered Autocorrelation Sequence for Robust Speech Recognition

الملخص EN

This paper presents a novel noise-robust featureextraction method for speech recognition using the robust perceptual minimum variance distortionless response (MVDR) spectrum of temporally filtered autocorrelation sequence.

The perceptualMVDR spectrum of the filtered short-time autocorrelationsequence can reduce the effects of residue of the nonstationaryadditive noise which remains after filtering the autocorrelation.

To achieve a more robust front-end, we also modify the robustdistortionless constraint of the MVDR spectral estimation methodvia revised weighting of the subband power spectrum valuesbased on the sub-band signal to noise ratios (SNRs), which adjustsit to the new proposed approach.

This new function allows thecomponents of the input signal at the frequencies least affected bynoise to pass with larger weights and attenuates more effectivelythe noisy and undesired components.

This modification resultsin reduction of the noise residuals of the estimated spectrumfrom the filtered autocorrelation sequence, thereby leading toa more robust algorithm.

Our proposed method, when evaluatedon Aurora 2 task for recognition purposes, outperformed all Mel frequency cepstral coefficients (MFCC) as the baseline, relative autocorrelation sequence MFCC (RAS-MFCC), and the MVDR-based features in several different noisy conditions.