Classification of AI- Generated vs. Human Voices Using Convolutional Neural Networks(CNNs) and Sprectrogram Analysis

The rapid development of AI technologies mainly in voice synthesis and cloning areas has caused considerable difficulties in distinguishing AI-generated and human voices. This brought huge risks to security, media authentications, and fraud detections. This paper presents a deep learning-based methodology for classifying AI-generated and human voices based on CNNs. We transform audio signals into their time-frequency representations, so-called spectrograms. The pre-trained CNN architectures to be fine-tuned for binary classification are VGG16 and ResNet50. The experiments show that the proposed models have high accuracy; ResNet50 outperforms VGG16 due to its deeper architecture. It provides a dogged means of differentiating AI-generated voices from human voices and gives another key tool in mitigating the risks connected to the malicious utilization of AI-driven voice technologies.