An Optimization-Driven Framework for Deep Learning-Based Spoken Language Recognition

This study presents a mathematical optimization framework for spoken language recognition (SLR) based on the integration of deep learning, metaheuristic optimization, and neuro-fuzzy classification components into a single recognition pipeline. The problem is formulated as high-dimensional nonlinear optimization task, where the objective is to maximize classification performance through optimal parameter selection and feature representation. Audio signals are transformed into spectrogram images, and discriminative features are extracted using an SE-DenseNet architecture. To address the challenges of hyperparameter tuning, the Golden Jackal Optimization (GJO) is employed to explore the search space and identify near-optimal configurations. The extracted features are subsequently classified using an adaptive neuro-fuzzy learning system (ANFLS), enabling improved decision boundaries under uncertainty. The proposed framework is evaluated on a dataset comprising three languages. Experimental results demonstrate an average accuracy of 99.26% and an F-score of 98.86% on the considered dataset, indicating strong in-domain performance compared with conventional machine learning and deep learning models. The results validate the effectiveness of the proposed optimization-driven formulation and highlight its potential for solving complex pattern recognition problems in high-dimensional spaces.