An Improved GA Based Modified Dynamic Neural Network for Cantonese-Digit Speech Recognition

An artificial Neural Network (ANN) is a well known universal approximator to model smooth and continuous functions (Brown & Harris, 1994). As ANNs can realize nonlinear models, they are flexible in modeling a wide variety of real-world complex applications, such as handwriting recognition, speech recognition, fault detection, medical inspection (Zhang, 2000), etc. ANNs being applied for pattern classification can be divided into two main categories: static and dynamic. Static pattern classification problems are usually tackled by multi-layer perceptron (MLP), radial basis feed-forward (RBF) networks and learning vector quantization (LVQ). However, limited by its structure of a traditional, a feed-forward network cannot model the correlation between the previous time frames and the current time frame. Thus, some dynamic applications, such as speech recognition, time varying prediction, dynamic control, etc. are difficult to be realized by static neural networks. Neither can feed-forward neural networks deal with problems without a fix dimension of input patterns. Recurrent Neural Network (RNN) (Engelbrecht, 2002; Kirschning et al., 1996; Zhang et al., 1993) and Time Delay Neural Network (TDNN) (Waibel et al., 1989) are used to overcome the limitations of feed-forward networks. They dynamically model time series cases; in other words, they are predictor networks that predict the next data frame from the current data frame. ANNs operate in two stages: learning and generalization. Learning of a neural network is to approximate the behavior of the training data while generalization is the ability to predict well beyond the training data (Zhang, 2000). In order to have a good learning and generalization ability, a good tuning algorithm is needed. In this chapter, Genetic Algorithm (GA) is used as the tuning algorithm for training neural networks. GA is a directed random search technique (Hanaki et al., 1999; Michalewicz, 1994; Pham & Karaboga, 2000) that is widely applied in optimization problems (Hanaki et al, 1999; Michalewicz, 1994; Pham & Karaboga, 2000). It is especially useful for complex optimization problems when the number of parameters is large and the analytical solutions are difficult to obtain. GA can help find out the globally optimal solution over a domain. It has been applied in different areas such as fuzzy control (Leung et al., 2004), path planning, modeling and classification (Setnes & Roubos, 2000), tuning parameters of neural/neuralfuzzy networks (Leung et al., 2003; Ling et al., 2003) etc. A lot of research efforts have been spent to improve the performance of GA. Different selection schemes and genetic operators