Improvement in Speech to Text for Bahasa Indonesia Through Homophone Impairment Training
JOURNAL OF COMPUTERS (TAIWAN) VOL.28, NO.5, 2017, PP.1-10. DOI:10.3966/199115992017102805001 ISSN: 19911599 (SCOPUS INDEX)
I.S. Areni, Indrabayu, A. Bustamin
In this research, an approach for increasing accuracy in speech to text application is conducted using Mel Frequency Cepstral Coefficient (MFCC) which then trained with Backpropagation Neural Network (BPNN). A Set of Bahasa Indonesia homophones data speech is used for training and validation. The record is taken from 6 native adults comprises of 3 males and 3 females. Working in 16 KHz sampling mode the data is stored in WAV format. A confusion matrix is used to validate system with and without homophone locking learning. A significant improvement are observed from the experiment. The percentage of accuracy is increased from 53.33 to 93.4 from male samples. From females records the increment is even higher. The accuracy percentage has risen from 36.8 to 93.33.
Keywords: Homophone, MFCC, BPNN, speech to text, confusion matrix