Sub-Word Based Amharic Word Wecognition : an Experiment using Hidden Markov Model (HMM)
No Thumbnail Available
Date
2002-06
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
In this study, the potential of Hidden Markov Model (HMM) for the development of
Amharic speech recognition system has been investigated and in the course of building
the recognizers the popular toolkit Hidden Markov Model Toolkit (HTK) was used.
In the process of building the recognizers, 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.
Since large vocabulary systems are envisaged, sub-word modeling is pursued. Sub-word
modeling refers to a technique whereby one HMM is constructed for each sub-word unit
(phoneme, triphone, syllable, etc.). Phonemes, tied-state triphones and CV-syllables have
been considered as the basic sub-word units and are used to build phoneme-based, tiedstate
triphone based and CV-syllable based recognizers respectively.
In this study, an extensible 170 word vocabulary is constructed and both speakerdependent
and speaker-independent models are built for 15 speakers (8 male and 7
female) in the age range of 20 to 30 using phonemes and tied-state triphones as the basic
units of recognition. Five untrained speakers who had no involvement in training the
models are also used to test the speaker-independent models.
The results obtained are promising and have shown the potential of tied-state triphones as
good sub-word units for Amharic. In fact, phonemes also have produced encouraging
recognition performance. Even though CV-syllables appear to be more convenient for
Amharic, this research has not proved that and is recommended for further research.
Description
Keywords
Word Recognition