DESIGNING AUTOMATIC SPEECH RECOGNITION FOR GE’EZ LANGUAGE

ABEBE, TSEGAYE

dc.contributor.author	ABEBE, TSEGAYE
dc.date.accessioned	2020-03-17T09:48:09Z
dc.date.available	2020-03-17T09:48:09Z
dc.date.issued	2020-03-17
dc.identifier.uri	http://hdl.handle.net/123456789/10545
dc.description.abstract	Language is a typical spiritual communication tool used by human beings in order to share or exchange their knowledge, skills, opinions wishes, commands, thanks, wisdoms and cultures to other people by speaking or using different ways. Currently, human beings are communicating with electronic devices using their speech with the help of automatic speech recognition system. Automatic speech recognition is the process of converting acoustic speech signals into its equivalent text form. Researches on automatic speech recognition have done on foreign or local languages Amharic. Amharic and Ge’ez languages have redundant letters with the same sound. Researches in Amharic speech recognition are conducted by normalization of the repeated letters. Though, those redundant letters have different usage and meaning in Ge’ez language. Ge’ez language is a classical language of our country which is looking to speech recognition research. Ge’ez language has its own letters and numbering system. However, some letters have lost their sound and they are making confusion during the formation of words in its writing system. The aim of this study is to investigate the possibility of developing automatic speech recognition for Ge’ez language. In this study hidden Markov modeling technique is applied using sphinx 4 trainer. Since there is no recorded or prepared Ge’ez corpus, the investigators developed both text and speech corpora and among the developed corpus 4818 sentences for training and 433 sentences for testing used by selecting seven speakers’ Ge’ez audio randomly. Two experiments were performed using two different language models and two testing methods (online and offline) were performed for evaluation of the system. Both experiment 1 and experiment 2 have shown 90.88% and 68.48%-word accuracy rate respectively by testing with sphinx tool and the average word accuracy is 79.70%. As well as for testing the system using the developed interface with the same testing data the word accuracy is 67.79%. Homophones and hetero-phones in Ge’ez are challenges for speech recognition. In order to increase the accuracy of recognizer, maximizing the size of corpora is the future direction. Phone based, syllable based and gemination were the other future works.	en_US
dc.language.iso	en	en_US
dc.subject	Computer Science	en_US
dc.title	DESIGNING AUTOMATIC SPEECH RECOGNITION FOR GE’EZ LANGUAGE	en_US
dc.type	Thesis	en_US