Analisis Perbandingan Ekstraksi Fitur Wav2vec2 dan MFCC terhadap Klasifikasi Pengucapan Huruf Iqra’ Menggunakan Algoritma Bagging

Abstract

The Bagging method suffers from suboptimal accuracy due to limitations in feature extraction and complex variations in audio data. Although Bagging is known to reduce overfitting and improve model stability (especially in algorithms such as Decision Tree), its performance is highly dependent on the quality of the features produced. The Wav2vec2, MFCC, and Wav2vec2-MFCC combination methods were used as methods for extracting audio features. The dataset ratio used is 80:20. This study formed three scenarios for each model (models with Wav2vec2, MFCC, and combination feature extraction). The first scenario uses primary data, the second scenario uses secondary data, and the third scenario uses combined data (primary and secondary). The results show that the best performance was achieved in the second scenario using the MFCC feature extraction method, achieving an accuracy of 82.74%, precision of 83.12%, recall of 82.64%, and an F1-score of 82.59%. These results indicate that the MFCC feature extraction method performs quite well when used in Iqra‟ pronunciation classification with the bagging method. The Iqra‟ letters with the lowest performance are the letters „da‟ and „ja‟ due to phonetic similarity, intonation variation, and noise interference. This study emphasises the importance of feature selection and balanced data distribution in Iqra‟ pronunciation classification, as well as the need to consider the diversity of the data used.