Syahroni Hidayat, Tatyantoro Andrasto, Febry Putra Rochim, Mulil Khaira, Muhammad Hilmy Herdiansyah, Faila Nadhifatul Aryza, Abdulloh
Congenital anomalies such as Cleft Lip and Palate (CLP) significantly affect speech production, necessitating early detection and intervention. This study proposes a speech recognition system for CLP patients by integrating Wavelet Transform and Mel Frequency Cepstral Coefficients (MFCC) features with a Long Short-Term Memory (LSTM) model. Several types of wavelets were evaluated through a 5-fold cross-validation strategy, focusing on performance metrics such as accuracy, sensitivity, and specificity. Results indicated that the Coif1 wavelet achieved the highest mean accuracy and sensitivity while maintaining stable performance across folds. Statistical validation using the Friedman test showed that although no metric reached statistical significance at the 0. 0 5 threshold, sensitivity differences approached significance, suggesting potential variability among wavelet types. Spearman correlation analysis further revealed strong positive relationships among accuracy, sensitivity, and specificity. The findings demonstrate that the combination of single-level wavelet-MFCC decomposition and LSTM architecture effectively captures the temporal dynamics of CLP speech patterns. Future work will explore additional wavelet families, introduce multi-level wavelet decomposition, and investigate alternative deep learning architectures to enhance system performance and robustness. © 2025 IEEE.
Universitas Negeri Semarang, Dept. of Electrical Engineering, Semarang City, Indonesia; Universitas Negeri Semarang, Dept. of Computer Engineering, Semarang City, Indonesia; Universitas Negeri Semarang, Dept. of Informatics and Computer Engineering Education, Semarang City, Indonesia
