Perception of band-specific speech quality distortions: Threshold and preference data and their objective prediction

The relative importance of different critical bands for the perception of s peech transmission quality was investigated. Two algorithms were introduced that generate a band-specific modulated-noise distortion in the speech sig nal. Detection thresholds were measured as a function of the center frequen cy of the band used fur generating the distortion. The pairwise speech qual ity preference was assessed at 4 different center frequencies for 3 differe nt levels of the modulation depth that were selected relative to the respec tive detection thresholds obtained From the detection experiment. The exper imental results were compared with predictions obtained from an objective s peech quality measure qc [1]. The detection thresholds were modeled by assu ming a constant value of qc at threshold. In order to predict subjects' thr esholds, a constant spectral weighting had to be applied to the critical ba nd filter channels on which the measure qc was based. This contrasts to the increasing weight with increasing frequency originally proposed for the ob jective measure qc. Also the pairwise speech quality preference ratings wer e modeled by the respective difference Delta qc with constant spectral weig hting of qc. The constant spectral weighting appears appropriate for detect ion and quality comparison tasks, whereas the absolute assessment of speech quality distortions of various codecs from codec selection test data bases appears to depend more strongly on high-frequency bands. That application therefore requires increasing weights for increasing center frequencies.