Learning individual talkers’ structural preferences