Search Results

The Rapid Progress Of Generative Models Has Enabled The Creation Of Highly Realistic Synthetic Voices, Commonly Known As Audio Deep Fakes. While These Technologies Have Beneficial Applications In Entertainment And Assistive Systems, They Also Pose Significant Risks In Misinformation, Fraud, And Security Breaches. Detecting Audio Deep Fakes Remains A Challenging Task Due To The Increasingly Natural Prosody, Timbre, And Linguistic Coherence Of Synthesized Speech. This Paper Proposes A Multi-stage Framework For Audio Deep Fake Detection That Integrates Complementary Strategies Across Acoustic, Linguistic, And Deep Feature Domains. In The First Stage, Handcrafted Acoustic Features Such As Mel-frequency Cepstral Coefficients (MFCCs) And Spectral Distortions Are Extracted To Capture Low-level Signal Artifacts. The Second Stage Leverages Linguistic Consistency Analysis To Identify Irregularities In Phoneme Duration And Speech Rhythm. Finally, Deep Learning–based Embeddings From Pre Trained Models Are Employed To Capture High-level Semantic And Prosodic Patterns. By Combining These Heterogeneous Feature Spaces Through Ensemble Classification, The Proposed Framework Achieves Robust Performance Against State-of-the-art Synthesis And Voice Conversion Systems. Experimental Results On Benchmark Datasets Demonstrate Improved Generalization Across Multiple Attack Scenarios, Highlighting The Potential Of The Framework As A Practical Tool For Safeguarding Digital Communications Against Audio Forgery.