Deepfake Voice Recognition 🎤🔍

Welcome to the Deepfake Voice Recognition App, a Streamlit-based web application designed to identify whether an audio file is a deepfake or a real voice. This project leverages machine learning models trained on the DEEP-VOICE dataset. The app is a casual exploration of using machine learning to distinguish between real and AI-generated speech.

Try the live app here: Deepfake Voice Recognition

---

🌟 Features

• Upload Audio: Users can upload audio files.
• Choose Model: Select between two pre-trained models:
  • Random Forest (rf_model.joblib)
  • LSTM (lstm_model.keras)
• Deepfake Detection: The app predicts whether the uploaded voice is real or fake (AI-generated).

---

📊 Dataset Overview

This project uses the DEEP-VOICE dataset, introduced in the study "Real-time Detection of AI-Generated Speech for DeepFake Voice Conversion" by Bird and Lotfi (2023). The dataset includes:

• Real Speech: Human voices recorded from eight well-known figures.
• Fake Speech: AI-generated voices created by converting one speaker's voice to another using Retrieval-based Voice Conversion (RVC).

Key Dataset Features:

• Raw Audio: Available in the REAL and FAKE directories.
• Pre-extracted Features: Stored in DATASET-balanced.csv, used for training the models in this project.

Ethical Considerations:

The dataset was developed to address the rising ethical concerns about generative AI in speech, such as privacy violations and voice misrepresentation. A successful detection system could notify users when AI-generated speech is detected in real-time scenarios like calls or conferences.

---

🛠️ Project Structure

deepfake-voice-recognition/
├── models/                                 # Trained models
│   ├── rf_model.joblib                     # Random Forest model
│   ├── lstm_model.keras                    # LSTM model
│   └── lstm_scaler.joblib                  # Pre-trained scaler object for feature scaling
├── notebooks/                              # Jupyter notebooks for training and experimentation
│   └── deepfake-voice-recognition.ipynb    # Random Forest and LSTM training notebook
├── app.py                                  # Streamlit app script
├── pyproject.toml                          # Poetry configuration
├── poetry.toml                             # Python dependencies
└── README.md                               # Project documentation

---

🚀 Getting Started

Prerequisites

• Python 3.8+
• Poetry (for managing dependencies)

Installation

1. Clone the repository:

   git clone https://github.com/malikfm/deepfake-voice-recognition.git
   cd deepfake-voice-recognition

2. Install dependencies using Poetry:

   poetry install

3. Run the Streamlit app locally:

   poetry run streamlit run app.py

4. Open the app in your browser: http://localhost:8501

---

🔍 Model Training

I utilized two distinct approaches for model training, i.e. Random Forest and LSTM. Detailed training processes and corresponding code can be found in the notebooks/ folder.

---

🔗 Live Demo

Experience the app live: Deepfake Voice Recognition

---

📜 License

This project is licensed under the MIT License. See the LICENSE file for details.

---

🎉 Acknowledgments

• Dataset: DEEP-VOICE
• Study: "Real-time Detection of AI-Generated Speech for DeepFake Voice Conversion" by Bird, J.J. and Lotfi, A. (2023).