In this paper, we propose an approach for determining significant differences in speech of typically developing children, children with Autism Spectrum Disorder (ASD) and Down syndrome. To start solving this problem, we performed an automatic graphemic and morphological analysis of transcribed children’s dialogues. Sixty-two children (20 children with typical development, 14 with Down syndrome, 28 with autism spectrum disorder) discussed standard set of questions with experimenters; for further analysis, only the children’s replicas were used. A total of 25 linguistic features were extracted from each dialogue: the number of replicas, the number of sentences, the number of tokens, the number of pauses, the number of unfinished words and the part of speech composition. To reduce the dimensionality, we performed Kruskal-Wallis tests to assess differences in these features among the studied groups of children, which allows to select 12 significant features. These features were incorporated into tree models such as Gradient Boosting, Random Forest, Ada Boost. All machine learning methods showed high performance, which allows to conclude about a good differentiating ability of features. Our best method showed a classification accuracy of 83%.