Detecting Deep Fake Voice using Machine Learning

Abstract

With the rapid advancement of deep learning techniques, the creation of synthetic media, particularly deep fake voices, has become increasingly sophisticated and accessible. This poses significant challenges in maintaining trust and authenticity in audio-based content. In response, this project proposes a machine learning-based approach to detect deep fake voices. The project begins by curating a diverse dataset consisting of genuine and deep fake voice samples, covering various demographics, accents, and emotional expressions. Pre-processing techniques are applied to clean and standardize the audio data, followed by feature extraction to capture relevant characteristics of the voice signals. For model development, a Convolutional Neural Network (CNN) architecture augmented with recurrent layers is employed, leveraging its ability to learn spatial and temporal features from the spectrogram representations of the audio. The model is trained on the prepared dataset using categorical cross-entropy loss and optimized through backpropagation.Evaluation of the trained model is conducted on a separate test set, measuring performance metrics such as accuracy, precision, recall, and F1-score. Post- processing methods, including thresholding and smoothing, are applied to refine the model’s predictions and enhance robustness. The proposed approach offers a promising framework for detecting deep fake voices in audio content, contributing to the ongoing efforts to combat the spread of misinformation and preserve the integrity of digital media. However, ongoing research and collaboration across disciplines are essential to address emerging challenges and ensure the responsible deployment of deep fake detection technologies.