Mrkomiljon/DeepVoiceGuard

DeepVoiceGuard: Real-Time Audio Authenticity Detection

DeepVoiceGuard is an advanced AI-powered tool for detecting whether an audio file is genuine or AI-generated. Built using RawNet-based architecture and trained on ASVspoof datasets, this model is optimized for real-time inference using ONNX format.

🚀 Features

• Real-Time Detection: Analyze audio files quickly and efficiently to determine authenticity.
• Sliding Window Processing: Processes long audio files in segments for accurate classification.
• ONNX Optimized: Faster inference compared to traditional formats.
• Interactive Demo: Test the model using our Streamlit application.

📚 Model Overview

• Architecture: RawNet-based Neural Network
• Frameworks Used: PyTorch, ONNX
• Dataset: Trained on ASVspoof 2019 Challenge dataset(LA)
• Classes:
  • Real: Genuine human speech
  • Fake: AI-generated or spoofed audio

🛠 Installation

Install the necessary dependencies:

pip install onnxruntime librosa numpy requests streamlit

🔧 How to Use Using the ONNX Model

import streamlit as st
import librosa
import numpy as np
import onnxruntime as ort
import os
import requests

# Audio padding function
def pad(x, max_len=64600):
    """
    Pad or trim an audio segment to a fixed length by repeating or slicing.
    """
    x_len = x.shape[0]
    if x_len >= max_len:
        return x[:max_len]  # Trim if longer
    # Repeat to fill max_len
    num_repeats = (max_len // x_len) + 1
    padded_x = np.tile(x, (1, num_repeats))[:, :max_len][0]
    return padded_x
# Preprocess audio for a single segment
def preprocess_audio_segment(segment, cut=64600):
    """
    Preprocess a single audio segment: pad or trim as required.
    """
    segment = pad(segment, max_len=cut)
    return np.expand_dims(np.array(segment, dtype=np.float32), axis=0)  # Add batch dimension

# Download ONNX model from Hugging Face
def download_model(url, local_path="RawNet_model.onnx"):
    """
    Download the ONNX model from a URL if it doesn't already exist locally.
    """
    if not os.path.exists(local_path):
        with st.spinner("Downloading ONNX model..."):
            response = requests.get(url)
            if response.status_code == 200:
                with open(local_path, "wb") as f:
                    f.write(response.content)
                st.success("Model downloaded successfully!")
            else:
                raise Exception("Failed to download ONNX model")
    return local_path
# Sliding window prediction function
def predict_with_sliding_window(audio_path, onnx_model_path, window_size=64600, step_size=64600, sample_rate=16000):
    """
    Use a sliding window to predict if the audio is real or fake over the entire audio.
    """
    # Load ONNX runtime session
    ort_session = ort.InferenceSession(onnx_model_path)

    # Load audio file
    waveform, _ = librosa.load(audio_path, sr=sample_rate)
    total_segments = []
    total_probabilities = []

    # Sliding window processing
    for start in range(0, len(waveform), step_size):
        end = start + window_size
        segment = waveform[start:end]

        # Preprocess the segment
        audio_tensor = preprocess_audio_segment(segment)

        # Perform inference
        inputs = {ort_session.get_inputs()[0].name: audio_tensor}
        outputs = ort_session.run(None, inputs)
        probabilities = np.exp(outputs[0])  # Softmax probabilities
        prediction = np.argmax(probabilities)

        # Store the results
        predicted_class = "Real" if prediction == 1 else "Fake"
        total_segments.append(predicted_class)
        total_probabilities.append(probabilities[0][prediction])

    # Final aggregation
    majority_class = max(set(total_segments), key=total_segments.count)  # Majority voting
    avg_probability = np.mean(total_probabilities) * 100  # Average probability in percentage

    return majority_class, avg_probability

# Example
result = predict("example.wav")
print(f"Prediction: {result}")

📊 Performance Metrics Equal Error Rate (EER): 4.21% Accuracy: 95.8% ROC-AUC: 0.986

🛡 License This project is licensed under the MIT License.

✉️ Contact For inquiries or support, please contact:

• GitHub: Mrkomiljon
• Hugging Face: DeepVoiceGuard