smpdshukla/deepfake-audio-detection-app

🎙 Deep-fake-Audio-Detection

Deep fake Audio Detection is AI/ML based project that uses classification models to detect whether audio is real or a machine generated. The pipeline includes:

1. 1.Audio Preprocessing: Converts raw audio into machine-readable features. It extracts Mel-Frequency Cepstral Coefficients (MFCCs), which help the model capture nuances in human speech.
2. 2.Feature Normalization: Applies a pre-trained scaler to normalize features across different audio samples, ensuring consistency and reducing model bias caused by amplitude variations or background noise.
3. 3.Model Prediction: Multiple deep learning models are used in parallel,CNN,RNN,BiLSTM,GRU,XGBoost for the prediction.
4. 4.Ensemble Learning: Outputs from all base models are combined using a Logistic Regression meta-classifier, which assigns optimal weights,and XGBoost, which captures complex non-linear relationships.
5. 5.Web App Integration: A user-friendly Streamlit interface allows non-technical users to upload audio files and instantly receive predictions.
6. 6.Performance: Deep learning models like CNN and BiLSTM excel individually (95% and 92% accuracy respectively), but combining them through stacking significantly improves robustness. Our final ensemble model achieved ~94% overall accuracy, outperforming individual models while remaining efficient enough for real-world deployment.

🚀 Features

1.Deepfake vs. Real Detection: Identifies whether an uploaded audio clip is genuine human speech or synthetically generated using deepfake technology.

1. 1.Advanced Ensemble Learning: Combines strengths of CNN, RNN, BiLSTM, GRU, and XGBoost to deliver high-accuracy predictions.
2. 2.Cross-Platform Deployment: Works seamlessly as a Streamlit interface, depending on the use case. 4.User-Friendly Interface: Non-technical users can easily upload audio files without requiring command-line interaction.
3. 3.Scalable and Extendable: The architecture is modular, allowing easy integration with new models, datasets, or cloud deployment platforms.

📦 Dataset

1. 1.Source: for-2seconds dataset taken from FoR dataset.(Publicly available speech datasets & synthetic deepfake audio.)

1. 1.Temporary Testing Dataset:

1. 1.Follow Path: temp/for-2seconds/testing/

1. 1.Contains short 2-second audio clips for quick evaluation.

1. 1.real/ → Human voice samples.

1. 1.fake/ → AI-generated voice samples.

1. 1.Supported Formats: .wav (default), .mp3 and .flac (when enabled in uploader).

📊 Models Used

✅ Meta-classifier combines all base models’ predictions for final classification.

🛠 Tech Stack & Tools

• —Languages: Python 3.9+
• —Deep Learning: TensorFlow, Keras
• —Machine Learning: XGBoost, Scikit-learn, Logistic Regression (meta-classifier)
• —Audio Processing: Librosa, SoundFile
• —Data Handling: NumPy, Pandas
• —Visualization: Matplotlib, Seaborn
• —UI & Deployment: Streamlit
• —Environment: Jupyter Notebook, Anaconda

▶️ Usage

1. 1.Install dependencies:

    pip install -r requirements.txt

1. 1.Download the dataset consisting of Fake and Real audio files for testing: 📂 Fake-or-Real (FoR) Dataset on Kaggle or from the temp folder from the github repo

1. 1.Run the Streamlit app:

    streamlit run app.py

4.Upload an audio file via the interface.

5.View prediction — whether the audio is Real or Deepfake.

📉 Visual Outputs in App

• —🎵 Waveform of uploaded audio
• —🎼 MFCC feature plot
• —📊 Stacked model performance chart
• —🗂 Final result in a highlight box – "Real" or "Fake" with confidence %

📈 Future Improvements

1.Support for multilingual audio.

2.Real-time streaming detection.

3.Cloud deployment for scalability.

4.Integration of Explainable AI (XAI) for transparency.