Figure 4.

(Color online) Comparison of recognition scores of the original (unmodified) and the three different modifications (feature-amplified, feature-attenuated and feature-removed) of the sounds f101ga (a) and m102da (b), to exemplify the modifications. When the burst is attenuated by 6 dB, the score shifts to the right, indicating that a higher SNR is required for the same performance as the unmodified sound. When the feature is amplified by 6 dB, the score shifts to the left, indicating improved performance by the magnitude of the shift. When the feature is completely removed, the scores drop more than when the feature is attenuated by 6 [dB] (dashed line), indicating the importance of the modification region. One standard-deviation bars are shown for f101ga (a) at a few SNRs. They are calculated using the equation $\sigma \left(P_c\right)=\pm \sqrt{P_c(1-P_c)/N}$. Only a few error bars are shown since they only depend on P_c.