Speaker Verification Model USING LST

Abstract: This project presents a speaker verification model using LSTM networks to detect the authenticity of voices, distinguishing between real and deepfake audio. With the growing threat of AI-generated voices being used for identity theft and misinformation, this model aims to address those risks. The system uses MFCCs (Mel-frequency cepstral coefficients) to capture key speech features and distinguish real voices from altered ones. The dataset includes both real and deepfake samples from well-known public figures like Joe Biden and Donald Trump, with techniques like time-stretching and pitch-shifting used to improve the model’s performance. Built with LSTM layers and trained for 50 epochs using the Adam optimizer, the model achieved an accuracy of 84.62%, proving effective in detecting fake voices. This LSTM-based model offers a valuable solution for enhancing voice authentication systems, especially in critical areas like security systems, online transactions, and biometric authentication. Its robust performance demonstrates the potential to counter the rising threat of deepfake technology