Predicting foreground SH, SL and BNH DAM scores for multidimensional objective measure of speech quality

Current objective measures of speech quality (Rix et al. (2000), Beerends et al. (1994) attempt to evaluate degraded speech by calculating a single distance measure between the original signal and the synthesized signal being evaluated. The distance measure is usually carried out after both the original and synthesized signal have been transformed to represent the effect of the auditory periphery. However, the fact that the subjective judgement of quality is based on a multidimensional perceptual space representation suggests that a measure that is based on predicting a multitude of independent perceptual characteristics, would yield better results and be applicable to a wider range of distortions and speech synthesis systems. This paper presents such a multidimensional approach to objective evaluation of speech quality and is directly motivated by the work of Voiers (2001) from which the subjective evaluation procedure known as diagnostic acceptability measure (DAM) was created. While the DAM is a subjective measure of the detectability of the distortions identified by Voiers, this work reports on the first baby steps taken for objective evaluation of a subset of those same parametric distortions determined to be the principal components of the quality space from a previous statistical analysis (Sen (2001)).