Modular neural networks for low-complex phoneme recognition

We present a Modular Neural Network (MNN) for phoneme recognition within the framework of a hybrid system (neural networks and HMMs) for speakerindependent single word recognition. With this approach, we are taking the computational effort into account which is used as an additional criterion for assessing the system performance. The main idea of the proposed MNN is the distribution of the complexity for the phoneme classification task on a set of modules. Each of these modules is a single neural network which is characterized by its high degree of specialization. The number of interfaces, and therewith the possibilities for infiltering external acoustic-phonetic knowledge, increases for a modular architecture. Moreover, after the development of a suitable topology for the MNN, each of the modules can be optimized for its specific phoneme recognition task. This is done by detecting and pruning irrelevant input parameters and leads to a more efficient system in terms of memory and computational requirements.