M. Shoaib, M. Awais, S. Masud, S. Shamail, and J. Akhtar (Pakistan)
Speech Recognition, Formants, Intensities, and Generalized Regression Neural Networks
This paper presents an application of concurrent generalized regression neural networks for Arabic speech recognition. It also discusses a comparative performance of concurrent as well as single generalized regression neural networks. The recognition system developed is based on 2240 speech samples, obtained from 80 different speakers. The sound is first sampled and then preprocessed to obtain formant frequencies and intensity contours. These formant frequencies and intensity contours are used to calculate the location, trend and gradient values. In order to automate the recognition process, these `location, trend and gradient' features are given as inputs to the concurrent generalized regression neural network and single generalized regression neural network. The validation tests of the concurrent neural networks showed 93.37% recognition accuracy whereas the single neural network demonstrated classification accuracy level up to 82.59%.
Important Links:
Go Back
