الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 209 |  |  |
| | | from | 209 | | |
Abstract
The ultimate speech recognition system must be free of the constraints of isolated words, small vocabulary, constrained grammar, and speaker dependence. In this thesis, proposed solutions for vocabulary size, continuous speech, and grammar problems are introduced. The computational linguistic techniques play an important role in relaxing most of these constraints. Also, the incorporation of the linguistic knowledge into the recognition algorithm increases the recognition accuracy significantly. The current proposed recognition algorithm is based on the strategy of analyzing and understanding the input utterance, rather than the incorporate template matching process.
It is worth nothing that the input utterance is treated first by the digital signal processing module. Thus , input to the proposed recognizer is a set of phonetic segments. Each segment is represented by a group of alternatives , rather than a signal unique one. Every alternative, has an attached score denoting its probability of occurrence.
A set of computational linguistic analysis techniques is applied to this phonetic sequence alternatives. The analysis techniques implemented in this research are the morphological, syntactic, and statistical ones, in addition to the construction of a lexical knowledge base to assist these analyzers in the recognition of the input utterance.
During these analyses , the input sequence of segment alternatives is filtered to get rid of the incorrect segment hypotheses.
The small vocabulary constraint is relaxed by using a lexical knowledge base instead of the usually used word vocabulary. This lexical knowledge base occupies a much smaller storage area than that of the word vocabulary. The grammar constraint is relaxed by using a simple (loose) grammar using the Definite Clause Grammar technique. The main advantage of this grammar is its capability to correlate with both the morphological analyzer and the lexical knowledge base. The implemented syntactic analyzer will get useful information from this correlation process.
The problem of continuous speech is solved by using the Finite State Machine approach . Adeterministic finite state machine is built for each analyzer. The machines are then integrated analyzer. These machines are then integrated to yield the recognizer machine. Anew proposed algorithm for building these machines is introduced to suit for the nature of the Arabic language. Also, the speed of the recognition process is enhanced to deterministic nature of these machines.
The proposed recognizer is quite reliable due to the independence of the recognition algorithm of the speech unit.
Thus, the recognition algorithm can adapt itself with any kind of speech unit without affecting the accuracy og the recognizer.
Finally, an application of recognizing the Arabic digits either isolated or connected, is introduced . An accuracy over 98 % is recorded for the recognition of isolated digits, and about 97 % for the connected digits. It is worth nothing that the implemented recognition algorithm for the previously used training for the isolated digits. This means that the same storage area could be used . Thus, the required storage area for the proposed recognition algorithm is extremely saved, in contrast to the large amount of storage needed by most of the current recognizers during the recognition of connected digits.
This application could be successfully implemented in the Direct Inward Dialing (DID) systems.
These systems simulate the job of the human telephone operator in a Private Branch Exchanges (PBX).