Abstract
Application of Neural Networks (NN) in speech recognition technology is equally widespread as it is efficient. In this project, NN is applied at preliminary level to develop a whole-word based speech recognition prototype system. It conducts on a specific alphabets ranged from /A/ to/J/. For simple recognition task like this, the system offers a platform towards further elaborated speech synthesis such as syllables or phoneme based recognition. Theoretically, Linear Predictive Coding with an order of 12 coefficients is derived during the feature extraction process. A typical network encompasses three layered Feed-Forward Multilayer Perceptron with Back Propagation learning algorithm is proposed for the classification phase. In order to attain a speaker independent recognition system, various training and testing data samples are collected from 10 speakers, consists of 5 males and 5 females. Concluding the final result, however, illustrates the difficulty in developing such system whereby the best recognition rate achieved was only at 76% with approximately 30% error rate for individual spoken words.
Website : http://ir.fsksm.utm.my/483/
No comments:
Post a Comment