The Role of Artificial Intelligence Algorithms in Detecting Audio Forgery and Promoting Sustainable Development: An Applied Study Using Zero-Frequency Filter and Neural Networks

Abstract

This study proposes a robust audio spoof detection framework by integrating Zero Frequency Filtering (ZFF) with a Convolutional Neural Network (CNN). In the preprocessing stage, ZFF is applied to remove DC offset, followed by the extraction of 13 Mel-Frequency Cepstral Coefficients (MFCCs) per frame. All audio samples are standardized to 3 seconds, resulting in 94 frames per sample. The dataset is split using a stratified approach into 60% training, 20% validation, and 20% testing to ensure balanced class distribution. The proposed CNN model achieves high classification accuracy of 99.05% on clean data, demonstrating strong capability in distinguishing between genuine and spoofed audio signals. Furthermore, the model maintains robust performance under real-world environmental noise conditions, including wind, rain, sirens, and engine sounds, even at a low Signal-to-Noise Ratio (SNR) of -10 dB. The results are consistent with previous studies that emphasize the effectiveness of combining signal processing techniques with deep learning models for noise-robust classification. The use of MFCC features and stratified data splitting contributes to improved evaluation reliability. However, the study is limited to the ASVspoof 2019 Logical Access dataset and does not include statistical significance testing, which may impact generalizability. Future work should evaluate the model on diverse datasets and incorporate additional performance metrics such as Equal Error Rate (EER) and F1-score.