Application of genetically optimized neural networks for hindi speech recognition system

Automatic speech recognition (ASR) can be formulated as statistical pattern classification problem. In this approach, normally short term features are derived from the speech signal at front-end and then evaluated at back-end using the hidden Markov models (HMMs) or artificial neural networks. In this paper, we present a novel approach by using multilayer perceptrons optimized with the help of genetic algorithm. A combination of both short term and long temporal context features has been used as a sequence of acoustic feature vectors. Experimental result shows significant improvement by using the proposed framework for spoken Hindi digit recognition in general field conditions as well as in noisy environment.