We describe here the similarities using sophisticated textual
descriptors (male voice, romantic piano, etc.). Of course these
descriptors are not to be deduced by the system – this task is
probably impossible to achieve – but are used only as comments to
a global similarity that is intrinsically unlabelled.