Hybrid Spectral and Statistical Feature Modelling with Cross-Correlation and Ensemble Learning for Robust Turkish Voice Command Classification

İkizler N.

Süleyman Demirel Üniversitesi Fen Bilimleri Enstitüsü Dergisi, cilt.30, sa.1, ss.67-82, 2026 (TRDizin)

Özet

Accurate classification of Turkish voice commands is essential for advancing voice-controlled technologies and enabling seamless human-computer interaction in native language contexts. This study systematically evaluates multiple feature extraction models capturing temporal, spectral, and time-frequency characteristics of speech signals to enhance recognition accuracy. Six feature vector models were developed, with the final model integrating Information Gain-based feature selection and Linear Predictive Coding-derived formant frequencies to create a comprehensive and discriminative representation. Classification was performed using six widely adopted algorithms: Random Forest, k-Nearest Neighbors, Multilayer Perceptron, Logistic Model Tree, Support Vector Machine, and an Ensemble voting method combining Random Forest, Multilayer Perceptron, and Logistic Model Tree. The Ensemble voting classifier demonstrated superior performance, achieving an accuracy of 93.94%, significantly outperforming individual classifiers and baseline models. This study contributes to the literature by presenting a robust, explainable, and high-performing feature framework tailored for Turkish voice command recognition. The integration of spectral, temporal, and articulatory features enables improved discrimination of speech commands, offering valuable insights for future voice-activated applications in Turkish language environments.