الفهرس Only 14 pages are availabe for public view
 |  | from | 165 |  |  |
| | | from | 165 | | |
Abstract
Current state of the art speech recognition systems are based on subword units as the basic building block. The accuracy of these systems depends mainly on two processes; the segmentation process which aimed to segment the speech signal into subword units, and the classification process in which the labeling of the
segmented subword units is occurred.
For the segmentation process a new algorithm is developed for segmenting the Arabic phonemes. The algorithm based on the wavelet transform along with the theory of fractal dimension and achieves a segmentation accuracy of 80.4% when applied on a sample database of 420 Arabic words articulated in modern standard Arabic.
For the classification process; two speech recognition techniques were applied; Hidden Markov Models (HMMs), and Support Vector Machines (SVMs).
Besides a new combined approach between Hidden Markov models and Support
Vector Machines is developed. This combined approach achieves about 10%
improvement relative to HMM based system or SVM based system.
For increasing the overall recognition accuracy,a phoneme sequence alignment
approach is done using Needleman-Wunsch algorithm, for comparing the resulting phoneme sequence with a dictionary of pronunciations,in order to approximately recover the intended word uttered.
This thesis presents a complete Arabic speech recognition system with a
comparable accuracy of the other speech recognition systems for other
languages. The overall recognition accuracy of the proposed Arabic speech
recognition system reached about 79%.