​You are cordially invited by the COE Dept., to attend a Guest Seminar, on the above given title, by Mr. Khalid Mohamed Oqlah Nahar, Full-Time PhD Student (Lecturer-B – COE), on Wednesday, May 15th, 2013, at 12:30 P.M., in Building 24, Room 115.​

Abstract: Recent research in speech recognition focuses on speaker independent continuous speech recognition. Continuous speech is more challenging because of variability in the words pronunciations, due to dialects, speaker age, gender, emotional status as well as the nearby words. Phones are used as sub-word units for pronunciation spelling. Each phone is represented by an acoustic model. The acoustic model produces the probability of a given feature vector for a particular phoneme. The acoustic units, used in representation of language words are called phones. The phone set in continuous speech recognition systems is usually selected by expert phoneticians. Some existing studies investigated the optimality of English phone set, while no study is being done to investigate the optimality of the Arabic phone set which is currently used in Arabic speech recognition. An alternative way for modelling pronunciation in terms of acoustic units is the automatic derivation of the sub-word units. Based on data-driven methods, this research work, aims at determining an inventory of sub-word units to be used in Arabic speech recognition. Additionally, a new dictionary will be produced based on the generated sub-words units using data-driven approach. A hybrid HMM/LVQ-ANN recognition methodology for the existing Arabic phones set will be conducted. This hybrid methodology will then be applied on the extracted sub-word units. In summary, the main goal of this work is to automatically generate a set of sub-word units that will improve the Arabic speech recognition performance, reduce the words error rate and enhance existing Arabic speech recognition system.

During this research, multiple clustering and segmentation techniques will be evaluated for the purpose of automatic extraction of Arabic sub-word units. The accuracy of the recognition system based on the newly extracted sub-words units will be measured and compared with the Automatic Speech Recognition (ASR) baseline.

Dr. M. El-Shafie, Professor, SE Department, is his thesis advisor.​