What is presented is a new approach for implementing Bosnian phoneme recognition. While most of the literature on phoneme recognition is based on hidden Markov models (HMM), or on the recognition by neural networks (NN) of one type, the present system is implemented by hybrid cascaded LVQ/Elman NN. This model was created, because we noted that some types of NN achieve better recognition rate for some phonemes, while the other types of NN better recognize other phonemes. Presented system uses LVQ NN as a front-end recognizer, and depending on the obtained output, makes the re-recognition of the same phoneme by Elman NN. This system achieved higher recognition accuracy then standalone NN models.Index Terms-Bosnian phoneme recognition, Melfrequency cepstral coefficients, hybrid cascaded LVQ/Elman neural network.