Usage Instructions
Download and Install:
Download the latest release
Extract the entirety of the release to a directory on your machine
Simply run the .exe file to launch the application.
Using the Application:
Open the executable file.
Use the graphical interface to upload audio files.
The system will process the audio and display whether it is genuine or spoofed.
Using the Library:
Download the .tar.gz file or input "pip install AVR" into your CMD
in a .py or .ipynb file add "import AVR"
now you can use AVR.query_function() with the path to your file as input. and the function will return the prediction value, True for a bona-fide voice sample, and False for a spoofed voice.
Remember to use AVR.flush() after use to free up the model memory from your system.
Overview
The AI Voice Recognize (AVR) project, developed by Guy Ben Ari and Ynon Friedman from Afeka College of Engineering, focuses on detecting spoofed audio files and distinguishing them from genuine human voices using advanced machine learning techniques. Dataset
ASVspoof2019 dataset is used, featuring a diverse range of real and spoofed audio files for model training and evaluation. Approach
Architecture: Utilizes LSTM networks for sequential audio data and Conv2D layers for feature extraction from MFCC images.
Model: Features are concatenated and classified using fully connected layers.
Development Environment
Prototyping and Training: Conducted using JupyterLab Notebook.
Version Control: Managed with Git.
Integration
The system is packaged as an executable file for ease of use, providing both real-time and batch processing capabilities. Testing Breakdown
Tests Conducted: Includes UI functionality, audio processing accuracy, and model integration.
Results: Includes test scripts, issues identified, and corrective actions taken.
Limitations and Solutions
Data Inconsistency: Resolved through normalization techniques.
Integration Issues: Fixed by integrating with HuggingFace.
Resource Constraints: Addressed by upgrading VRAM.
Performance: Enhanced with CUDA.
UI and Security: Improved based on feedback and added encryption.
Contact
For inquiries or collaboration:
Guy Ben Ari: [email protected]
Ynon Friedman: [email protected]