Abstract
Conference Title: 2016 International Conference on Engineering & MIS (ICEMIS) Conference Start Date: 2016, Sept. 22 Conference End Date: 2016, Sept. 24 Conference Location: Agadir, Morocco The creation of a robust system for speech recognition requires a great effort and resources. One of the crucial elements in the creation of such systems is the phonetic dictionary. The creation of such a dictionary requires phonologists experts and linguists to create the best possible dictionary. But the Arabic phonologists have two theory about the Arabic phonemes. The first one says that vowels phonemes are independent of consonants phonemes. The second one says that vowels have no phonemes but only consonants have phonemes and the vowels are a part of them. In this paper, we have created two phonetic dictionaries. Each dictionary is based on one of these two theories. Then these two dictionaries have been tested by two Arabic automatic speech recognition systems. The results show that the dictionaries built based on the first theory are much better and give a word error rate less than other dictionary build based on the second theory. In our case, we got an average of Word Error Rate of 13.41% using the first dictionary and 23.09% using the second dictionary. Our new study, and for the first time, settle the debate between phonologists about the relationship between vowels phonemes and consonants phonemes. Also our new study shows that the construction of an Arabic automatic speech recognition system will be useful only if the phonetic dictionary takes in consideration the independence between the vowels phonemes and the consonants phonemes. As our study indicates the weakness of HMM at the representation of the variability in length of vowels.