Discriminant learning for hybrid HMM/MLP speech recognition system using a fuzzy genetic clustering

We suggest for this study a fuzzy-genetic process for speech clustering, in the framework where the result of fuzzy c-means (FCM) clustering was used as initial population for genetic algorithms (GA). The approach is used in a hybrid HMM/ANN system using an Artificial Neural Network (ANN) to compute the observation probabilities in the states of the Hidden Markov Models (HMM). Experimental results obtained with continuous databases of various sizes in two languages (Arabic and French) show a significantly improved recognition accuracy with respect to the discrete HMM and regular hybrid HMM/ANN model using traditional clustering approaches.