IMPLEMENTASI MEL-FREQUENCY CEPSTRAL COEFFICIENTS DAN CONVOLUTIONAL NEURAL NETWORK UNTUK PENGENALAN HURUF HIRAGANA

Abstract

Japanese language learning has gained increasing interest in Indonesia; however, learners often experience difficulties in mastering Hiragana characters due to their large number and phonetic similarities. Speech recognition technology can be utilized as a supportive learning medium, particularly for improving pronunciation and enhancing learners’ understanding of Hiragana characters. This study aims to develop a Hiragana speech recognition system based on Mel-Frequency Cepstral Coefficients (MFCC) for feature extraction and Convolutional Neural Networks (CNN) for classification. The dataset consists of 46 Hiragana characters, with each character recorded 20 times by four speakers, resulting in a total of 3,680 audio samples. The research stages include audio signal preprocessing, MFCC feature extraction, data augmentation, CNN model training, and performance evaluation using classification metrics. Experimental results indicate that the proposed model achieves an accuracy of 95% on the test data, with most Hiragana characters being correctly recognized. Misclassifications mainly occur among characters with similar phonetic characteristics. These results demonstrate that the MFCC-based CNN approach is effective for Hiragana speech recognition and has potential to be applied as an interactive digital learning medium for Japanese language education.