Hidden Markov Model Based Large Vocabulary, Speaker Independent, Continuous Amharic Speech Recognition
No Thumbnail Available
Date
2003-06
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
This study investigated the possibility of developing large vocabulary, speaker
independent and continuous speech recognizer for Amharic language. The recognizer was
developed using Hidden Markov Model; and the Hidden Markov Modeling Toolkit was
used to implement it.
In the process, a corpus developed by Solomon Tefera was used to get the required data
for training and testing the models. It was a database comprised of 8000 utterances that
were used for training and 500 plus sentences for development and evaluation. The data
was preprocessed in line with the requirements of the HTK toolkit. In order to support the
acoustic models, a bigram language model was constructed. In addition, pronunciation
dictionary was prepared and used as an input.
Since the recognizer was meant to recognize large vocabulary and continuous speech,
phonemes were opted 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
was built, it was promoted to triphone based system in which the left and right contexts
were considered and modeled. Besides, the mixture components of the states of the models
were incremented in view of optimizing the performance of the recognizers.
Description
Keywords
Speech Recognition