Evaluation of prosodic quality is always a challenging task dueto the nature of prosody with multiple form-function valid pro-files. When voice of people with Down syndrome (DS) is ana-lyzed, diversity increases making the problem even more chal-lenging. This work is framed in our activities for developinglearning games for training oral communications of people withintellectual disabilities. In this context automatic evaluationof prosodic quality is a must for deciding whether game usersshould repeat activities or continue playing and to inform ther-apists about the particular difficulties of users. In this paper wepresent a procedure for the selection of informative prosodicfeatures based on both the distance between human rated rightand wrong productions and the distance with respect to produc-tions of typical users. It is new with respect to previous worksthe use of mixed models to rate the importance of taking into ac-count the type of activity and speaker for estimating the qualityof the prosodic productions.