audio-deepfake-detection

Urdu Deepfake Audio Detection Models

This repository contains trained models for detecting deepfake audio in Urdu language.

Models Included

• SVM: Support Vector Machine with RBF kernel
• Logistic Regression: Binary classification model
• Perceptron: Single-layer perceptron
• DNN: Deep Neural Network with 2 hidden layers

Features Used

• MFCC (Mel-frequency cepstral coefficients)
• Mel Spectrograms
• Chroma features
• Spectral contrast
• Zero-crossing rate

Performance

Usage

import pickle
import librosa
import numpy as np

# Load model and scaler
with open('audio_svm_model.pkl', 'rb') as f:
    model = pickle.load(f)
with open('audio_scaler.pkl', 'rb') as f:
    scaler = pickle.load(f)

# Load and process audio
audio, sr = librosa.load('audio_file.wav', sr=16000)
features = extract_features(audio, sr)  # Use feature extraction function
features_scaled = scaler.transform(features.reshape(1, -1))

# Predict
prediction = model.predict(features_scaled)
probability = model.predict_proba(features_scaled)

Citation

If you use these models, please cite the original dataset:

@dataset{csalt_urdu_deepfake,
  title={CSALT Urdu Deepfake Detection Dataset},
  author={CSALT},
  year={2024}
}