MARGIN–RELIABILITY GATED ENSEMBLE FOR ROBUST SPEECH ACCENT RECOGNITION USING HANDCRAFTED ACOUSTIC FEATURES

Abstract

Speech accent recognition is an emerging area of speech technology with applications in human–computer interaction, automated transcription, and language learning systems. Although deep learning has recently dominated the field, conventional machine learning models based on handcrafted acoustic features remain highly relevant due to their interpretability, lower computational cost, and suitability in resource-constrained environments. This paper proposes the Margin–Reliability Gated Ensemble (MR-GE), a novel machine learning framework that integrates calibrated margins from multiple baseline classifiers (Support Vector Machines, Random Forests, k-Nearest Neighbors, and Gaussian Mixture Models), accent-aware reliability weights derived from per-class validation profiles, and sparse non-negative feature-group gating across MFCC, PLP, prosodic, and spectral descriptors. The framework is systematically evaluated against baseline classifiers and optimized through dimensionality reduction and feature selection methods such as PCA, LDA, mRMR, and ReliefF. Experimental results on multi-accent speech datasets demonstrate that MR-GE achieves significant improvements in recognition accuracy, F1-score, and calibration reliability compared to individual classifiers. These findings establish a robust baseline-to-optimized machine learning pipeline for accent recognition and provide a foundation for future integration with deep learning approaches.