New Features Using Robust MVDR Spectrum of Filtered Autocorrelation Sequence for Robust Speech Recognition
Figure 4
Comparison of the proposed as the gain controlling the steepness of the weighting function and the resulted weighting function with the one used for extracting robust features when no RAS filtering is used (such as those suggested for RPMCC features). The SNRs in the figures refer to estimated subband SNRs at the output of Mel filter banks followed by equal loudness curve and power law of hearing, as shown in Figure 3(b) and, therefore, are different from SNRs estimated over the whole signal.
