This project delves into the challenging domain of synthetic speech detection, employing a sophisticated analysis framework that integrates short-term and long-term prediction traces. Leveraging cutting-edge deep learning methodologies such as Convolutional Neural Networks (CNN), Wave Net, an undisclosed model denoted as ISTM, and Recurrent Neural Networks (RNN), our methodology endeavors to establish a robust mechanism for identifying synthetic speech across diverse contexts. Central to our approach is the extraction of a comprehensive array of features, encompassing both short-term attributes like Zero Crossing Rate and Spectral Control, and long-term characteristics such as Mel-Frequency Cepstral Coefficients (MFCC) and Chroma features. Through meticulous data preprocessing, model training, and rigorous evaluation, our aim is to construct a system that exhibits high accuracy in discerning synthetic speech instances. This endeavor not only contributes significantly to the advancement of speech processing techniques but also holds promise for real-world applications in fraud detection, voice authentication, and content verification. By addressing the burgeoning challenge of synthetic speech detection, our project endeavors to pave the way for enhanced security measures and trustworthiness in voice-based systems and applications.

Keywords: ISTM, Synthetic speech, RNN,