Multidimensional evaluation and predicting overall speech quality

The quality of speech samples has been traditionally evaluated in subjective listening tests using 5-point Absolute Category Rating (ACR) scales in Listening Only Tests (LOT) as recommended in ITU-T P.800 [1]. Those tests provide the listening quality aspect of speech quality. There are other tests are under discussion and proposed in order to assess in detail individual perceptual dimensions of speech. In this paper we investigate the relationship between the overall listening quality obtained in an ITU-T P.800 ACR subjective test and the rating of the same signals in four dimensions proposed by Wältermann [2], namely noisiness, discontinuity, coloration and loudness. The database we use is composed of conditions and speech signals extracted from an ACR LOT used in the ITU-T P.863 evaluation, processed by simulated and live telecommunication channels [3]. The signals have been re-scored using the four mentioned scales and are foreseen as contribution to the ITU-T P.AMD project. This paper focuses on the modeling of an ACR LOT score based on individual dimensional ratings under the assumption of orthogonality of the four dimensions.