Self-organizing map and multilayer perceptron for malay speech recognition

• Main Author: Goh, Kia Eng
• Format: Thesis
• Language: English
• Published: 2006
• Subjects: QA76 Computer software
• Online Access: http://eprints.utm.my/id/eprint/6385/1/GohKiaEngMFSKSM2006.pdf; http://eprints.utm.my/id/eprint/6385/; http://dms.library.utm.my:8080/vital/access/manager/Repository/vital:62257

Various studies have been done in this field of speech recognition using various techniques such as Dynamic Time Warping (DTW), Hidden Markov Model (HMM) and Artificial Neural Network (ANN) in order to obtain the best and suitable model for speech recognition system. Every model has its drawbacks and weaknesses. Multilayer Perceptron (MLP) is a popular ANN for pattern recognition especially in speech recognition because of its non-linearity, ability to learn, robustness and ability to generalize. However, MLP has difficulties when dealing with temporal information as it needs input pattern of fixed length. With that in mind, this research focuses on finding a hybrid model/approach which combines Self-Organizing Map (SOM) and Multilayer Perceptron (MLP) to overcome as well as reduce the drawbacks. A hybrid-based neural network model has been developed to speech recognition in Malay language. In the proposed model, a 2D SOM is used as a sequential mapping function in order to transform the acoustic vector sequences of speech signal into binary matrix which performs dimensionality reduction. The idea of the approach is accumulating the winner nodes of an utterance into a binary matrix where the winner node is scaled as value “1� and others as value “0�. As a result, a binary matrix is formed which represents the content of an utterance. Then, MLP is used to classify the binary matrix to which each word corresponds to. The conventional model (MLP only) and the proposed model (SOM and MLP) were tested for digit recognition (“satu� to “sembilan�) and word recognition (30 selected Malay words) to find out the recognition accuracy using different values of parameters (cepstral order, dimension of SOM, hidden node number and learning rate). Both of the models were also tested using two types of classification: syllable classification and word classification. Finally, comparison and discussion was made between conventional and proposed model based on their recognition accuracy. The experimental results showed that the proposed model achieved higher accuracy.