|M.Sc Student||Bedrik Larissa|
|Subject||Classification of Acoustic Characteristics|
of Speech in Hebrew and English
|Department||Department of Biomedical Engineering||Supervisors||PROFESSOR EMERITUS Isak Gath|
|PROFESSOR EMERITUS Hillel Pratt|
Accent is one of the most important factors (next to gender) that influence speaker-independent recognition algorithms. Most recognition algorithms are gender dependent, with no accent information taken into account. By extracting the accent information from the speech signal and using this information for an accent dependent recognition system, machine speech recognition can be improved. The research work deals with English accent of Hebrew language and with Hebrew accent of English language. The subject of the research is the question of characterization of a speaker's second language, already at the early stage of pronunciation. The purpose is to find out how speech characteristics are changed and which of them have the greatest meaning in the determination of accents. The acoustic differences between English and Hebrew words, spoken as first and second language, were measured for vowels and consonants. The research deals with the vowels (/a/, /e/) and consonants /k/, /p/, /t/). The database was obtained from one bilingual speaker. The following features were examined: formant values at the beginning of each word, formant values in the central part of a vowel, and consonant and vowel durations. After the features were obtained, a logistic regression algorithm was run to classify the received data. Initial classification results of English as a mother tongue for the bilingual speaker case was 91.7% correct. The classification results of Hebrew as a mother tongue under similar conditions were 79.5% correct. This work is only an initial study and there is a place for further research to improve the classification results.