Toward Robust Audio Deepfake Detection in Real Acoustic Environments: Limitations of MFCC-Based Classical Models and the Advantages of Deep Learning with Domain Adaptation
Year: 2026
Type: conference-paper
Abstract: The proliferation of AI-driven voice synthesis tools has raised serious concerns about audio-based biometric authentication. This paper examines the practical limitations of classical machine learning models built on 26-dimensional MFCC (Mel-frequency cepstral coefficient) feature vectors when deployed in real acoustic environments. While such models achieve competitive equal error rates on clean benchmarks like ASVspoof (Automatic Speaker Verification Spoofing), they exhibit substantially elevated false positive rates when tested against genuine voice recordings captured in home settings with natural background noise, reverberation, and compression artifacts. Through a structured user survey involving 130 participants, we document that public concern about voice cloning and audio deepfakes is both widespread and practically grounded. We argue that classical models overfit dataset-specific noise signatures rather than learning synthesis-intrinsic artifacts. To address this, we propose a CNN (convolutional neural network) based deep learning framework operating on mel spectrograms, augmented with domain adaptation, as a substantially more robust alternative. Experimental results confirm that this approach reduces false positive rates on real-world audio while maintaining strong detection performance on synthetic samples.
Institutions: Kazakh-British Technical University, Astana IT University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203614569