Audio Deepfake Detection With Machine Learning Classifiers and Engineered Acoustic Features

Abstract

The voice synthesis systems powered by artificial intelligence (AI) can now synthesize speech that can be perceived as similar to that of a person, which makes it increasingly difficult to verify authenticity. The increasing realism of synthetic speech has created a need for reliable detection methods that can distinguish genuine and manipulated audio while maintaining interpretability and practical applicability. This article proposes a deepfake audio detection framework based on engineered acoustic features and machine learning classifiers to enhance the detection of deepfake audio. The proposed system makes use of openSMILE ComParE_2016 to create 6373 detailed spectral, prosodic, and voice-quality features, and XGBoost feature selection is used to select the 20 most discriminative features. This study focuses on reducing the high dimensionality of acoustic representations while retaining the most relevant information required for effective classification. XGBoost, LightGBM, random forest, multilayer perceptron (MLP), support vector machine (SVM), and logistic regression were six classifiers tested on a balanced corpus with real and synthetic audio. An XGBoost model had an accuracy of 97.85% and F1-score of 0.952 on unseen data, demonstrating the effectiveness of the selected feature subset under the evaluated experimental conditions. The findings show that interpretable and high-performance deepfake detection can be achieved with a cautious choice of acoustic features, along with effective classifiers, which will prove useful in trustworthy digital media usage and forensic practice.