Deepfake Audio Detection for Urdu Language Using Deep Neural Networks

Audio Deepfakes, which are highly realistic fake audio recordings driven by AI tools that clone human voices, recent advancements in Text-to-Speech (TTS) and Vocal Conversion (VC) technologies have made it easier to create realistic synthetic and imitative speech, making audio Deepfakes a common and potentially dangerous form of deception. Well-known people, such as politicians and celebrities, are often targeted. They are tricked into saying controversial things in fake recordings, making them a nuisance on social media. Even kid’s voices are cloned to scam parents into ransom payments, etc. Therefore, developing effective algorithms to distinguish fake audio from real audio is critical to preventing such frauds. Various machine learning (ML) and deep learning (DL) techniques have been created to identify Deepfake audios. However, most of these solutions are trained on datasets in English, Portuguese, French, and Spanish, expressing concerns about their correctness for other languages. The main goal of the research presented in this paper is to evaluate the effectiveness of deep learning neural networks in detecting Deepfake audios in the Urdu language. Since there was no suitable dataset of Urdu audio available for this purpose, we created our own dataset (URFV) using both genuine and fake audio recordings. The original/real Urdu audio recordings were gathered from random YouTube podcasts and generated as Deepfake audios using the RVC (Retrieve based voice conversion) model. Our dataset comprises three versions containing audio clips of 5, 10, and 15 seconds in duration. Each version includes samples from a total of 125 speakers (95 males and 30 females), and the data is divided into 70% for training, 15% for validation, and 15% for testing. We have built various deep learning neural networks like (RNN+LSTM, CNN+attention, TCN, CNN+RNN) to detect Deepfake audio made through imitation or synthetic techniques. The proposed approach extracts Mel frequency-cepstral-coefficients (MFCC) features from the audios in the dataset. When tested and evaluated, our model’s accuracy across datasets was noteworthy. 97.78% (5s), 98.89% (10s), and 98.33% (15s) were remarkable results for the RNN+LSTM model. In comparison, the CNN + RNN hybrid model equaled 92. 2% (5 s) and scored 98. 89% (10 s) and 93. 33% (15 s), while the CNN + Attention model recorded 98. 89% (5 s), 92. 22% (10 s), and 93. 33% (15 s). Despite its lower effectiveness, the TCN model obtained 70% (5s), 83.33% (10s), and 77.78% (15s).