Pengenalan Suara Vokal Bahasa Indonesia dengan Jaringan Saraf Tiruan Menggunakan Ciri Transformasi Wavelet Diskrit

Abstract

Vowel recognition is the main topic in speech recognition. There are six
Indonesian vowels, i.e. /a/, /i/, /u/, /e/, /ə/ and /o/. Feature extraction is an important
step in recognition system because the recognition rate depends on feature
extraction results. Vowel feature extraction via discrete wavelet transform (DWT) is
presented here. Mother wavelet db4 and sym4 are used. Minimum, maximum, mean
and standard deviation value of wavelet coefficients are extracted as vowel features.
DWT with level 2 decomposition obtains 12 features, level 4 decomposition obtains
20 features and level 6 decomposition obtains 28 features. Then, those vowel
features are used as an input of artificial neural network (ANN) with 2 hidden layers.
First hidden layer has 10 neurons and second hidden layer has variety 5 and 7
neurons. Backpropagation method is used to train the ANN. The vowel signals are
recorded from 10 female respondens and 10 male respondens. Each responden
pronounces six Indonesian vowels and syllable /ka/, /ki/, /ku/, /ke/, /kə/ and /ko/.
Experimental results show that the best recognition rate for the vowel is 85%, which
is obtained by using mother wavelet sym4, level 6 decomposition and 7 neurons for
second hidden layer, and the best recognition rate for the syllable is 80%, which is
obtained by using mother wavelet db4, level 6 decomposition and 5 neurons for
second hidden layer.