Audio Deepfake Detection using Enhanced CNN and Multi-Scale Ensemble Techniques

The rapid advancement of text-to-speech and voice conversion technologies has posed significant challenges to audio authenticity verification. This paper presents a novel approach for audio deepfake detection using a dual-model ensemble system combining Enhanced Convolutional Neural Networks (CNN) with Multi-Scale feature analysis. Our methodology leverages multi-feature extraction including spectral analysis, Mel-frequency cepstral coefficients (MFCC), and phase information to achieve superior detection accuracy. The Enhanced CNN model achieves 99.81% accuracy while the Multi-Scale Ensemble provides confidence scores ranging from 92–98% with robust generalization capabilities. Evaluated on the ElevenLabs dataset containing 2,561 audio samples (736 real, 1,825 fake), our system demonstrates state-of-the-art performance with real-time inference capabilities. The proposed architecture addresses critical challenges in audio deepfake detection including feature representation, model calibration, and computational efficiency, making it suitable for practical deployment in real-world scenarios.