Modeling of Integral Quality Based on Perceptual Dimensions - A Framework for a New Instrumental Speech-Quality Measure

In this contribution, the general framework for a new instrumental measure for end-to-end speech transmission quality is described. It is based on the notion that integral quality can be described by the global perceptual dimensions "discontinuity", "noisiness", and "coloration". The dimensions were identified through multidimensional analyses of telephone speech quality in an end-to-end context. Corresponding dimension impairment factors are defined which quantify the quality impairment of each dimension. Three additional Multidimensional Scaling experiments were conducted covering effects found in each of the global dimensions. By means of the resulting sub-dimensions, the global dimensions can be described in more detail. On the basis of the framework, signal-based or parametric dimension estimators are developed that predict either the quality impairment due to each dimension (i.e., the dimension impairment factors), or the underlying sub-dimension scores. It is shown that integral quality can be estimated by a combination of the dimension impairment factors, yielding a covered variance of R2 =90% of the auditory test data for a wide range of perceptually different conditions. Examples are given on how a particular dimension impairment factor can be predicted both on the basis of global dimensions and of sub-dimensions.