Towards a new reference system for subjective evaluation of coding techniques

This paper aims to produce a reference system that can simulate and calibrate degradations of conversational codecs which are currently used on telecommunications networks, for subjective assessment tests of voice quality. At first, 20 wideband codecs are evaluated through subjective tests in order to produce the multidimensional perceptual space underlying the perception of current degradations. Then, a verbalization task helps to identify attribute for each dimension: clear/muffle, background noise, noise on speech and hiss. These dimensions are characterized with correlates such as spectral centroid, energy in the silent part in the high frequency sub-band, ratio of brightness between deterministic part and residual part of the signal and correlation coefficient. This 4-dimension perceptual space is stable for male and female talkers. A new reference system is proposed. The phase of validation leads to a perceptual space with five dimensions including the four previous ones.