Accessing Gender Bias in Speech Processing Using Machine Learning and Deep Learning with Gender Balanced Audio Deepfake Dataset

The rapid advancement of AI has led to the rise of Audio Deepfakes (AD), which pose serious ethical and security concerns by accurately mimicking human speech. This research addresses the urgent need for effective AD detection, with a focus on gender bias that can reduce the effectiveness of detection models. We examined how gender affects the performance of both Machine Learning (Support Vector Machine, Random Forest, Logistic Regression, XGBoost) and Deep Learning (Deep Neural Networks, Convolutional Neural Networks) models using the GBAD dataset. Our findings show that models trained on female audio outperform those trained on male audio, likely due to the expressive nature of female voice features and high-pitched artifacts in FAKE audio. This highlights the need for more robust, gender-sensitive detection systems. Future work should focus on developing adaptive models to reduce gender bias, improving security, and creating lightweight models for wider public use.