SPEECH RECOGNITION OF KV-PATTERNED INDONESIAN SYLLABLE USING MFCC, WAVELET AND HMM

Abstract

The Indonesian language is an agglutinative language which has complex suffixes and affixes
attached on its root. For this reason there is a high possibility to recognize Indonesian speech
based on its syllables. The syllable-based Indonesian speech recognition could reduce the
database and recognize new Indonesian vocabularies which evolve as the result of language
development. MFCC and WPT daubechies 3rd
(DB3) and 7th
(DB7) order methods are used in
feature extraction process and HMM with Euclidean distance probability is applied for
classification. The results shows that the best recognition rateis 75% and 70.8% for MFCC
and WPT method respectively, which come from the testing using training data test.
Meanwhile, for testing using external data test WPT method excel the MFCC method, where
the best recognition rate is 53.1% for WPT and 47% for MFCC. For MFCC the accuracy
increased asthe data length and the frame length increased. In WPT, the increase in accuracy
is influenced by the length of data, type of the wavelet and decomposition level. It is also found
that as the variation of state increased the recognition for both methods decreased.