Perceived quality evaluation : an application to sound reproduction over headphones

Quality evaluation of sound reproduction systems is ultimately a perceptual matter and is intuitively associated to the hedonic domain with questions regarding the goodness of a given system. However, beyond this integrative evaluation approach relying on a measure of overall impression, quality can also be tackled from a more elemental and analytical viewpoint with the aim of understanding the perceptual components affecting quality. The present thesis work focuses on this topic referred to as 'sensory analysis' and explores more specifically two different types of sensory techniques using quantitative scales associated to verbal descriptors to characterize stimuli: the consensus vocabulary and individual vocabulary methods. Sound reproduction over headphones is the application considered in this investigation on sensory analysis. Advanced applications for headphone rendering of spatial or three-dimensional sound and virtual surround sound have been developed but the perceived quality evaluation of such systems remains an issue. In the present work, the specific case of spatial enhancement algorithms for the reproduction of music or movie sound over headphones is addressed. Firstly, the development of a consensus vocabulary of headphone sound perception is reported in which a panel of screened assessors created a consensual set of sixteen attributes with their associated definition, word anchors and audio exemplars to describe the perceptual characteristics of a wide range of headphone sound stimuli. The applicability of this vocabulary is tested on two sets of headphone sound samples. Secondly, a procedure entitled 'individual vocabulary profiling' developed for rapid sensory evaluation of audio applications by inexperienced assessors using an individual elicitation method is presented and its application is illustrated on a set of spatial enhancement systems. Lastly, these two approaches are compared to highlight their respective benefits, challenges and limitations and their suitability for different sensory evaluation projects. Throughout this thesis, statistical techniques for the analysis of sensory data are also investigated. The attribute rating data resulting from the two sensory profiling experiments is analyzed systematically. Several advanced multivariate data analysis methods are explored and applied to these data sets including generalized Procrustes analysis, multiple factor analysis, hierarchical multiple factor analysis and parallel factor analysis. The concepts and techniques presented in this thesis are applicable to many audio quality problems and can be adapted to other perceptual domains