Audio Deepfake Detection

📌 Overview

This repository contains the implementation of an Audio Deepfake Detection model using XGBoost Classifier on extracted audio features. The project aims to detect AI-generated human speech by leveraging spectral and temporal audio characteristics.

Workflow (Simplified in 5 Steps)

1️⃣ Convert FLAC to WAV

• The dataset is in FLAC format, so we first convert it to WAV for easier processing.

2️⃣ Preprocess the Audio

• Extract important features like , spectrograms to understand the sound patterns.

3️⃣ Train the Machine Learning Model

• Use models like Transformers, XGBoost to learn differences between real and fake audio.

4️⃣ Evaluate the Model

• Test the trained model with new audio files and measure accuracy using confusion matrix, precision, and recall.

5️⃣ Predict and Detect Deepfakes

• The model analyzes a new audio file and predicts whether it is real or AI-generated (fake).

📂 Dataset Acquisition

The dataset used in this project is DEEP-VOICE, which consists of AI-generated and real human speech samples. Additional datasets were referenced from:

• ASVspoof 5

• flac_D of 6.6gb

  1️⃣ Handling Unlabeled Data – Since flac_D lacked labels, we analyzed metadata, extracted features, and used clustering techniques to separate real and fake audio samples.

  2️⃣ Preprocessing & Labeling – Some files were manually inspected, while feature-based classification helped in labeling data for effective model training.

🚀 Project Structure

Audio_Deepfake_Detection/
│── data/                        # Dataset storage
│   ├── flac_D/                  # Raw FLAC files
│   ├── wav_D_random10k/         # Converted WAV files
│   ├── features.csv             # Extracted audio features (from Excel file)
│   ├── mel_spectrograms.npy      # Precomputed spectrograms
│   
│
├── xgboost_deepfake_detector.pkl  # Trained XGBoost model
│
├── Audio Deepfake Detection Take.ipynb  # Main notebook for research & implementation
│
│── requirements.txt              # Dependencies list

Setup Instructions

1. Clone the Repository

First, download the project from GitHub:

git clone https://github.com/A-A-D-I-C-O-D-E/Audio-Deepfake-Detection.git
cd Audio-Deepfake-Detection

2. Create a Virtual Environment (Recommended)

To avoid conflicts, create and activate a virtual environment:

python -m venv venv
source venv/bin/activate  # On macOS/Linux
venv\Scripts\activate  # On Windows

3. Install Required Dependencies

Make sure you have all necessary libraries installed:

pip install -r requirements.txt

4. Run the Jupyter Notebook

Start Jupyter Notebook to train and test the deepfake detection model:

jupyter notebook

Then, open the notebook file and follow the steps inside.

📝 References

• Audio Deepfake Detection Papers
• Librosa Documentation
• XGBoost Documentation