AI Voice Detection System

A state-of-the-art deep learning system for detecting AI-generated voice content across multiple Indian languages. Built using Facebook's Wav2Vec2-XLSR-53 model with fine-tuning for multilingual voice authentication.

Overview

This project addresses the growing challenge of deepfake audio detection by providing a robust solution capable of distinguishing between human and AI-generated speech in five major Indian languages: Tamil, English, Hindi, Malayalam, and Telugu.

Key Features

• Multilingual Support: Supports 5 Indian languages (Tamil, English, Hindi, Malayalam, Telugu)
• High Accuracy: Fine-tuned Wav2Vec2-XLSR-53 model with advanced training strategies
• Real-time Inference: FastAPI-based REST API for production deployment
• Docker Ready: Containerized deployment with Docker support
• Audio Augmentation: Robust training with noise injection, pitch shifting, and time stretching
• Chunked Processing: Handles variable-length audio files efficiently
• Confidence Scoring: Provides detailed confidence metrics and explanations

Architecture

AI Voice Detection System
├── Training Pipeline
│   ├── train_wav2vec2.py    # Main training script
│   └── encode.py            # Audio encoding utilities
└── Production API
    ├── ai-voice-container/
    │   ├── app.py           # FastAPI inference server
    │   ├── Dockerfile       # Container configuration
    │   └── requirements.txt # Dependencies

Quick Start

Prerequisites

• Python 3.10+
• PyTorch 1.12+
• CUDA (optional, for GPU acceleration)
• Docker (for containerized deployment)

Installation

1. Clone the repository

git clone <repository-url>
cd AI-voice-detection

2. Install dependencies

pip install -r ai-voice-container/requirements.txt

3. Download the pre-trained model

# The model will be automatically downloaded from Hugging Face Hub
# Model: kripasree/ai-voice-detector

Running the API Server

Option 1: Direct Python Execution

cd ai-voice-container
uvicorn app:app --host 0.0.0.0 --port 8000

Option 2: Docker Deployment

cd ai-voice-container
docker build -t ai-voice-detector .
docker run -p 8000:8000 ai-voice-detector

Model Performance

Training Configuration

• Base Model: facebook/wav2vec2-large-xLSR-53
• Training Strategy: Two-phase fine-tuning (frozen + unfrozen)
• Batch Size: 1 (with gradient accumulation of 8)
• Learning Rate: 2e-6 (initial), 1e-6 (fine-tuning)
• Epochs: 7 (initial) + 2 (fine-tuning)
• Audio Length: Up to 7 seconds per sample
• Sampling Rate: 16kHz

Data Augmentation Techniques

• Random gain adjustment (0.8x - 1.2x)
• Noise injection (Gaussian noise, σ=0.003)
• Pitch shifting (±2 semitones)
• Time stretching (0.9x - 1.1x speed)

API Usage

Endpoint

POST /
Content-Type: application/json

Request Format

{
  "language": "english",
  "audioBase64": "base64_encoded_audio_data"
}

Response Format

{
  "status": "success",
  "language": "english",
  "classification": "AI_GENERATED",
  "confidenceScore": 0.8945,
  "explanation": "Synthetic patterns detected"
}

Classification Threshold

• AI Detection Threshold: 0.75
• Output Classes: HUMAN or AI_GENERATED
• Confidence Range: 0.0 - 1.0

Training Pipeline

Dataset Structure

HCL-DATA/
├── human/
│   ├── tamil/clips/
│   ├── english/clips/
│   ├── hindi/clips/
│   ├── malayalam/clips/
│   └── telugu/clips/
└── ai_generated/
    ├── tamil/
    ├── english/
    ├── hindi/
    ├── malayalam/
    └── telugu/

Training Process

1. Data Collection: Balanced sampling (1500 samples per language per class)
2. Preprocessing: Audio loading, cropping, and augmentation
3. Feature Extraction: Wav2Vec2 feature extraction at 16kHz
4. Model Training: Two-phase fine-tuning approach
5. Evaluation: 90/10 train-test split with epoch-wise evaluation

Running Training

python train_wav2vec2.py

Note: Update the BASE path in train_wav2vec2.py to point to your dataset location.

Development

Project Structure

• train_wav2vec2.py: Complete training pipeline with data augmentation
• encode.py: Audio encoding utilities for base64 conversion
• ai-voice-container/app.py: Production-ready FastAPI inference server
• ai-voice-container/Dockerfile: Container configuration for deployment

Key Dependencies

• torch: Deep learning framework
• transformers: Hugging Face transformers library
• librosa: Audio processing and analysis
• fastapi: Modern web framework for APIs
• uvicorn: ASGI server for FastAPI
• numpy: Numerical computing
• datasets: Hugging Face datasets library

Performance Metrics

The model achieves competitive performance across all supported languages with:

• High Precision: Minimal false positives in AI detection
• Robust Generalization: Handles various audio qualities and conditions
• Real-time Processing: Sub-second inference for typical audio clips
• Scalable Architecture: Designed for production workloads

Security & Ethics

• Privacy: No audio data is stored permanently
• Bias Mitigation: Balanced dataset across languages and demographics
• Responsible AI: Designed for legitimate content verification purposes

Contributing

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes (git commit -m 'Add amazing feature')
4. Push to the branch (git push origin feature/amazing-feature)
5. Open a Pull Request

License

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

Acknowledgments

• Facebook AI for the Wav2Vec2-XLSR-53 model
• Hugging Face for the transformers library and model hub
• HCL for providing the multilingual voice dataset

Contact

For questions, suggestions, or collaborations, please reach out through the project issues or contact channels.

---

Disclaimer: This tool is designed for legitimate content verification and research purposes. Users are responsible for complying with applicable laws and regulations when using this system.