| Résumé |
This paper reports on the analysis of a free classification of vocal imitations of everyday sounds. The goal is to highlight the acoustical properties that have allowed the listeners to classify these imitations into categories that are closely related to the categories of the imitated sound sources. We present several specific techniques that have been developed to this end. First, the descriptions provided by the participants suggest that they have used different kinds of similarities to group together the imitations. A method to assess the individual strategies is therefore proposed and allows to detect an outlier participant. Second, the participants’ classifications are submitted to a hierarchical clustering analysis, and clusters are created using the inconsistency coefficient, rather than the height of fusion. The relevance of the clusters is discussed and seven of them are chosen for further analysis. These clusters are predicted perfectly with a few pertinent acoustic descriptors, and using very simple binary decision rules. This suggests that the acoustic similarities overlap with the similarities used by the participants to perform the classification. However, several issues need to be considered to extend these results to the imitated sounds. |
