Dimension-based Speech-Quality Assessment: Instrumental Measure for the Overall Quality of Telephone-band Speech

This paper presents an instrumental quality measure for telephone-band speech that works on the basis of three quality dimensions: "directness/frequency content", "continuity", and "noisiness". The assessment of the single dimensions allows for an analysis of a system's speech transmission quality, while a suitable combination of the dimension ratings yields an estimate for its overall speech-quality. In this paper, the estimators for the dimensions underlying the overall-quality measure are only outlined, while the focus is on the assessment of the overall quality of telephone-band speech. The dimension-based quality measure presented here is explicitly trained on assessing the overall quality of speech signals, while it is only implicitly trained on measuring the three above-named quality dimensions. The MOS estimates show a correlation of r ≈ 0.85 with the MOS values obtained in auditory tests, whereas the resulting dimension estimates show a correlation of rho = 0.71. . .0.79 with the dimension ratings determined in auditory tests.