Audiolyt: A Lightweight Real-Time Framework for Deepfake and Auto-Tuned Audio Detection

The increasing prevalence of synthetic and autotuned speech has introduced significant challenges in audio authenticity verification. This paper presents Audiolyt, a lightweight, real-time framework for detecting deepfake and auto-tuned audio samples. Leveraging handcrafted audio features such as MFCC, Chroma, and Spectral Contrast, and a densely connected neural network (DNN), Audiolyt achieves a classification accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 6. 8 3 \%}$</tex>. The system is trained on a composite dataset combining ASVspoof 2019, SceneFake, and custom-generated autotuned samples using legacy songs and Audacity. Audiolyt is deployed as a web application using Flask and Render, offering real-time inference with minimal latency. Experimental results show that the DNN model outperforms CNN and BiLSTM-based architectures while maintaining low computational overhead, making Audiolyt suitable for deployment in constrained or embedded environments.