In this study, we present the results of classification experiments of induced dog barks in different contexts of behaviour. We applied four validation schemes to trained models in order to determine the level of individuals dependency for context classification. We did an analysis based on feature selection techniques to determine the best acoustic low-level descriptors for this task. Results showed that classification performance decreases when the model is evaluated leaving out acoustic information of individuals in the training stage. The acoustic feature set used in our experiments shown better results in comparison with other works using the same data.