Enhanced Modified BSD (EMBSD)

EMBSD [155] metric is based on the BSD algorithm. The EMBSD algorithm consists of a perceptual transform followed by a distance measure that incorporates a cognition model. The original and received signals are normalized and divided into 50% overlapping frames of 320 samples. Frames that exceed an energy threshold are transformed to the Bark frequency domain. The Bark coefficients are transformed to dB to model perceived loudness, then scaled. The first 15 Bark spectral components in each frame are used to compute the loudness difference. Only distortion that exceeds a noise masking threshold and lasts longer than 200 ms (10 frames) is considered. The distortion is computed as the average distortion for all valid frames. Both the original and distorted signals should be strictly synchronized, otherwise the performance becomes poor. Tests showed that its correlation with subjective results is relatively good for encoding impairments, however it is not able to evaluate the quality when network impairments are applied as we will show in Section 5.5.