AudioGuard: Deep Learning Based Telugu DeepFake Audio Detection
Siwani Karna, Samudrala Sai Santhoshi Haneesha, Poluru Reddy Jahanve, Peeta Basa Pati, Roshni M Balakrishnan · 2024
Recent advancements in deep fake audio technology have led to the creation of highly realistic synthetic voices that mimic human speech patterns. However, this technology raises ethical concerns due to its potential for misuse, including spreading misinformation, identity theft, and undermining trust in audio evidences. To address these issues, this research focuses on developing a method to detect deep fake audio specifically in the Telugu language. This research utilizes various Convolutional Neural Network (CNN) architectures like MobileNetV2, ResNet50, VGG16, and AlexNet to enhance accuracy and resilience. MobileNet achieved the highest accuracy of $\mathbf{9 9. 5 \%}$ and was chosen for deployment. The model is integrated into a user-friendly website using Flask, HTML, CSS, JavaScript, and MongoDB. This web application serves as a comprehensive solution for detecting Telugu deep fake audio, prioritizing user security, ease of use, and accuracy.