Browsing by Author "Abera, Hafte"

Now showing 1 - 2 of 2
  • No Thumbnail Available
    Item
    Hidden Markov Mode l Based Tigrigna Speech Recognition
    (Addis Ababa University, 2009-11) Abera, Hafte; Meshesha, Million (PhD)
    Conventional method of assessing the performance of gas turbine engine involves analyzing different engine parameters manually and comparing them with their respective acceptable limits in engines maintenance manuals . The method takes lengthy and complicated processes that demand personnel with many years of experience and allows subjective judgment of the personnel involved in evaluation process. Technological advances in design and constructions of gas turbine engines adds more eng in e parameters thereby setting more hurdles on the process of engines performance evaluation . This paper reports on the finding of a research that had the objective to build a model that classify the performance, either accepting or rejecting, of PT6A-27 model turboprop pas turbine engine . The engines cons id erred were those undergo ing evaluation for performance after they went through repair or overhaul. The data used to build the mode first passed through different data preprocessing and analyzing techniques . the model employed neural network built using back propagation algorithm on a neural network tool box found in MATLAB 6.5 . The model built classifies gas turbine engines by their performances into their respective classes . The classification accuracy found was encouraging for the model to be adapted in real problem. The outcome of this research can also put a corner stone for further researches in using data mining technique for gas turbine engine maintenance .
  • No Thumbnail Available
    Item
    Hidden Markov Model Based Tigrigna Speech Recognition
    (Addis Ababa University, 2009-11) Abera, Hafte; Meshesha, Million (PhD)
    This study aims to design speaker independent continuous Tigrigna recognition system. Tigrigna is a very productive language in terms of word forms because of its agglutinative nature. All work is done in HTK (Hidden Markov Model Toolkit) environment, except parsing and network transforming which utilizes perl programming language. In the process of building the recognizer, the speech data is recorded at a sampling rate of 16KHz and the recorded speech is then converted into Mel Frequency Cepstral Coefficient (MFCC) vectors for further analysis and processing. A corpus has been developed to get the required data for training and testing the models. The corpus is a database comprised of 250 utterances that are used for training and 50 sentences for testing and evaluation. The data is preprocessed in line with the requirements of the HTK toolkit. In order to support the acoustic models, a bigram language model is constructed. In addition, pronunciation dictionary is prepared and used as an input. Since the recognizer is designed to recognize continuous speech, Phonemes are used as the basic unit of recognition. However phonemes are known to be context independent units, given that the environment in which a sound is put makes a difference in the way it is pronounced. Thus, after the monophone based speech recognizer is built, it is promoted to triphone based system in which the left and right contexts are considered and modeled. The speech recognizer is then evaluated using the test dataset Performance result shows 60.32% word level correctness, 58.38% word accuracy, and 20 % sentence level correctness are obtained. The results are encouraging and with more optimization works better results can be achieved. To this ends further research works are recommended in line with the analysis and finding of this study. Key words: Speech Recognition, HMM, HMM based speech Recognition, Language Modeling.