In this work, we present a novel approach to one of computational paralinguistic tasks -automatic detection of deceptive and truthful information in human's speech. This task belongs to the aspects of destructive behaviour and was first presented at the International INTERSPEECH Computational Paralinguistics Challenge ComParE in 2016. The need of contactless method for deception detection follows from the fact that existing contact-based approaches such as polygraphs and lie detectors have multiple restrictions, which significantly limit their usage. Both for training and testing of the proposed models we used two English-language corpora (Deceptive Speech Database and Real-Life Trial Deception Detection Dataset). We extracted tree sets of acoustic features from those audio samples using openSMILE toolkit. The proposed approach includes preprocessing of the extracted acoustic features with the usage of methods for data augmentation and dimensionality reduction of feature space. We have got 1680 speech utterances and 986-dimensional informative feature vector for each utterance. The main part of the proposed approach is two-level recognition model, where the first level includes three models of gradient boosting (Catboost, XGBoost and LightGBM). The second level consists of logistic regression-based model for final prediction on truthfulness or deceptiveness that takes into account predictions from the first level. Using this approach, we have achieved the result of classification in terms of F-score = 85.6%. The proposed approach can be used both independently and as a component of multimodal systems for detection of deceptive and truthful utterances in speech, as well as in systems for detection of a destructive behaviour.