Deepfake Audio Detection Using MFCC-Based Acoustic Feature Extraction and Machine Learning Classification: A Comprehensive Framework

Abstract

Voice has always been one of the more trusted ways to identify people; we can identify our colleagues, family, and public figures by their voices. Meanwhile, that trust is being systematically exploited. The state of neural speech synthesis today means that AI-generated speech can sound convincing in casual conversations and, as in a handful of cited cases, has been used for financial fraud and targeted deception. In this paper, a detection system using Mel-Frequency Cepstral Coefficients (MFCCs) and a Support Vector Machine (SVM) classifier is presented to classify audio clips as real or generated. The system does not only use the direct MFCC values, but it also computes five other complementary spectral features: spectral centroid, spectral roll-off, zero crossing rate, RMS energy and spectral bandwidth, from which it forms a 45-dimensional feature vector. Noise is filtered, loudness normalised, duration trimmed, and silence standardised via audio pre-processing before any features are computed. The trained classifier is encapsulated within a Flask web app that accepts uploaded audio files, performs inference, and outputs predictions in real time. Results on a balanced benchmark test set of 4,000 clips yielded 95.2% accuracy, 94.1% precision, 95.8% recall, and 94.9% F1-score. These numbers outperform several CNN and TCN baselines in the recent literature at a significantly reduced computational cost. The paper suggests that when executed conscientiously, principled feature engineering can be a viable alternative to large-scale deep learning for this class of problems.