Over the last years, studies have highlighted many features liable to be used for the characterization of sounds by several elicitation methods. These various experiments have resulted in the production of a long list of sound attributes. But, as their respective meaning and weight are not alike for assessors and listeners, the analysis of the results of a listening test based on sound criteria remains complex and difficult. The experiments reported in this paper were aimed at shortening the list of attributes by clustering them in sound families from the results of two semantic tests based on either a free categorization (i) or use of a multi-dimensional scaling method (ii).