Abstract
This paper describes a new algorithm for estimating speech loudness function (dependency of each frequency component on total speech loudness) which is suitable when applied to objective rating assessments on loudness for telephoneconnections. This algorithm also provides a more efficient and reliable calculation method than the previous one. Loudness differences between reference system (NOSFER) and telephone systems are calculated from a 2-step model which consists of : i) non-linear transformation from sensitivity differences to psychometric differences and ii) weighted accumulation of psychometric differences. Model parameters are determined so as to minimize modified PSS (Prediction Sum of Squares), a criterion for selecting meaningful variables. Loudness balancing tests on various filters are carried out to examine this algorithm. Filters employed are implemented as FIR digital filters on a mini-computer software. Experimental results show that a few independent parameters can provide good agreement between computed and measured values and a parameter for nonlinearity has approximately the same value as the conventional power index of loudness.
