Compact modular neural networks in a hybrid speaker-independent speech recognition system

In recent years, the computational effort for novel speech recognition systems has increased much more than the resulting recognition rates. Therefore, we present an approach for overcoming this drawback by using a modular phoneme recognition system which is included in a hybrid system with a discrete hidden Markov model (HMM). The development of a suitable topology for the modular architecture and the determination of relevant input parameters for the modules are the essential aspects of this paper. The main idea of the proposed system is the distribution of the complexity for the classification task on a set of modules with a higher degree of specialization. Therewith, acoustic-phonetic knowledge can be selectively incorporated because of the higher number of interfaces. Important system features are module-specific selection of input parameters according to the decision tasks and the utilization of time delay neural networks (TDNNs) as well as static neural networks without time processing. The intention is to improve the recognition rate for speaker-independent phoneme recognition and, at the same time, to reduce the necessary effort for simulating the system after the initial learning phase