Measuring Normalizing Blocks (MNB)

MNB was developed at the US Department of Commerce in 1997 [139,140,141]. It emphasizes the important role of the cognition module for estimating speech quality. It models human judgment on speech quality with two types of hierarchical structures. It has showed to be relatively robust over an extensive number of different speech data sets. MNB transforms speech signals into an approximate loudness domain through frequency warping and logarithmic scaling. It assumes that these two factors play the most important role in modeling human auditory response. The algorithm generates an approximated loudness vector for each frame. MNB considers human listener's sensitivity to the distribution of distortion, so it uses hierarchical structures that work from larger time and frequency scales to smaller time and frequency scales. MNB employs two types of calculations in deriving a quality estimate: time measuring normalizing blocks (TMNB) and frequency measuring normalizing blocks (FMNB). Each TMNB integrates over frequency scales and measures differences over time intervals while the FMNB integrates over time intervals and measures differences over frequency scales. After calculating 11 or 12 MNBs, these MNBs are linearly combined to estimate the overall speech distortion. The weights for each MNB are optimized with a training data set.