RealTalk: Tagalog Deepfake Audio Detection

RealTalk is a deep learning project designed to detect Tagalog deepfake audio using a Vision Transformer (ViT) model.
The system converts audio clips into log-mel spectrograms, which are then analyzed by the model to classify whether the audio is real or fake.

This model was developed as part of my undergraduate thesis. It represents the backend detection engine; the web interface for uploading and analyzing audio is maintained in a separate repository.

Features

• Detects Tagalog deepfake audio clips
• Converts audio to log-mel spectrograms
• Uses a Vision Transformer (ViT) for classification
• Lightweight and easy to run locally

How It Works

1. Preprocessing (https://raw.githubusercontent.com/Anthunii/RealTalk_Inference_Model/main/__pycache__/RealTalk_Inference_Model_3.7.zip) – Input audio is converted into a log-mel spectrogram
2. Model Inference (https://raw.githubusercontent.com/Anthunii/RealTalk_Inference_Model/main/__pycache__/RealTalk_Inference_Model_3.7.zip) – The spectrogram is passed through the Vision Transformer
3. Classification (https://raw.githubusercontent.com/Anthunii/RealTalk_Inference_Model/main/__pycache__/RealTalk_Inference_Model_3.7.zip) – The model outputs whether the audio is Real or Deepfake (to be deployed)

Setup & Installation

pip install -r https://raw.githubusercontent.com/Anthunii/RealTalk_Inference_Model/main/__pycache__/RealTalk_Inference_Model_3.7.zip

Limitations

• Performance decreases on extremely noisy audio
• Primarily trained on Tagalog, limiting generalization to other languages
• Works best on clean, conversational voice clips

Recommendations / Future Work

• Expand dataset with more diverse speakers and recording environments
• Add more deepfake generation methods to improve robustness since AI keeps advancing
• Improve real-time inference performance
• Add explainability tools (e.g., attention visualization)
• Evaluate performance on larger benchmark datasets