Automatic Speech Liveness Detection

This paper addresses the growing challenge of detecting deepfake audio, which can mimic or alter speech using advanced machine learning techniques like GAN sand RNNs. Such audio manipulation poses risks in biometric security and fraud. The study focuses on developing detection methods using acoustic features such as Mel-Frequency Cepstral Coefficients (MFCCs) and zero-crossing rate (ZCR). Two neural network models were tested, with the best achieving accuracy in identifying synthetic audio. The results highlight the need for advanced detection algorithms to keep pace with evolving deepfake technologies, enhancing biometric and digital security systems.