Abstract
This document
represents and English translation of the original Persian Abstract
of Ms. Nosratighods Masters Degree Thesis entitled "Classification
of Persian Stop Consonants Based on Their Articulation Place via Wavelet
Transformation" which she completed while studying at Tehran
Polytechnic University and Artificial Intelligence Research Institute
in Tehran, Iran between January 2000 and June 2002
The most challenging phonemes in automatic speech recognition are stop
consonants.
So, a great deal of effort has been made to seek invariant acoustic
cues which help detecting the stop consonants, they include voicing,
manner and place of the articulation.
In this paper, we classified the Persian stop consonant s/ch, b, d,
g, q, j, p, t, and k/ via their voicing and place of articulation.
To achieve this goal, our classification was done in different steps;
the best features such as formants, discrete wavelet coefficients and
mel cepsrtum coefficients are extracted in each step on the FARSDAT
CV syllables.
Three different classifiers were used to classify Persian stop consonants,
LDA, QDA (Bayesian Classifiers) and MLP (Neural Classifier).
First, voiced and unvoiced stop consonants were separated via discrete
wavelet transform that resulted in 99.2% correct classification.
In the next step, /j/ and /ch/ were separated from voiced and unvoiced
plosives via VOT, voiced onset time.
Then, the glottal stop /?/ was separated from voiced stop consonants
/b,d,g,q/ via the invariability of the energy normalization in consecutive
frames in time domain.
Finally, voiced stop consonant and unvoiced stop consonants were classified
in their groups utilizing the four Stevens [] features, one of them
was related to formant measurements and the others were three relative
spectral amplitude of the burst, that resulted in high percent correct
classifications (about 96.4% for voiced and 98.5% for unvoiced).
As a result, the voiced and unvoiced stop consonants have been classified
correctly 89% and 95.6%, consecutively.
In all the steps, MLP was the best classifier.
Keywords: Automatic Speech Recognition, Wavelet, Formants, Place of
Articulation, Burst Spectrum