Multiple Pronunciations Modeling of Speaker Independent, Continuous Speech Recognition for Afaan Oromoo
No Thumbnail Available
Date
2013-01
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Automatic speech recognition, especially speaker independent continuous speech recognition, is
characterized by great variability in word pronunciation, including many variants that differ
grossly from dictionary prototypes. This is one factor in the poor performance of automatic
speech recognizers on speaker independent speech recognition.
One approach to handling this variation consists of expanding the dictionary with phonetic
substitution, insertion, and deletion rules. This study aimed at modeling multiple pronunciations
in speaker independent continuous speech recognition for Afaan Oromoo to handle
pronunciation variation. Hidden Markov Model and the Hidden Markov Modeling Toolkit were
used to implement it. For developing model for the language under study, a corpus containing
754 sentences collected from Bariisaa news paper and Afaan Oromoo bible (New Testament)
was used. The data collected was preprocessed in line with the requirements of HTK. Phonemes
were taken as base unit for recognition. Knowledge based pronunciation variation modeling
technique was used for modeling words with multiple pronunciations. Thus two models were
developed; one with canonical pronunciation and the other with alternate pronunciation to
compare their relative performance.
Accordingly, the performance achieved using canonical pronunciation was 81.09% and 83.82 %
correct for sentences and words respectively with word accuracy of 80.91 % while the
performance of alternate pronunciation was 83 .08% and 85. 11% sentences and words correct
respectively with 82.52% word accuracy.
Description
Keywords
Modeling, Speaker