Deepfake audio detection: a deep learning based solution for group conversations

Abstract

The recent advancements in deep learning and other related technologies have led to improvements in various areas such as computer vision, bio-informatics, and speech recognition etc. This research mainly focuses on a problem with synthetic speech and speaker diarization. The developments in audio have resulted in deep learning models capable of replicating natural-sounding voice also known as text-to-speech (TTS) systems. This technology could be manipulated for malicious purposes such as deepfakes, impersonation, or spoofing attacks. We propose a system that has the capability of distinguishing between real and synthetic speech in group conversations.We built Deep Neural Network models and integrated them into a single solution using different datasets, including but not limited to Urban-Sound8K (5.6GB), Conversational (12.2GB), AMI-Corpus (5GB), and FakeOrReal (4GB). Our proposed approach consists of four main components. The speech-denoising component cleans and preprocesses the audio using Multilayer- Perceptron and Convolutional Neural Network architectures, with 93% and 94% accuracies accordingly. The speaker diarization was implemented using two different approaches, Natural Language Processing for text conversion with 93% accuracy and Recurrent Neural Network model for speaker labeling with 80% accuracy and 0.52 Diarization-Error-Rate. The final component distinguishes between real and fake audio using a CNN architecture with 94 % accuracy. With these findings, this research will contribute immensely to the domain of speech analysis.