TrueTone AI Voice & Deepfake Detector API

A lightweight, high-performance FastAPI service for detecting AI-generated, synthetic, and deepfake audio with state-of-the-art deep learning models.

Designed for real-time fraud detection, voice authentication, and security research use cases.

---

Key Features

• Detects AI-generated vs Human speech\
• High-performance PyTorch inference pipeline\
• Advanced signal preprocessing & augmentation\
• Secure API with API Key authentication\
• Dynamic model loading from Hugging Face\
• Minimal dependencies --- no FFmpeg or heavy audio stacks\
• Production-ready FastAPI architecture

---

Core Architecture

The system combines a self-supervised speech representation model with a graph-attention anti-spoofing classifier.

Feature Extractor Backbone

• Facebook MMS-300M (Wav2Vec2) model
• Provides rich multilingual speech embeddings
• Robust against noise and channel variations

Backend Classifier

• Custom AASIST (Audio Anti-Spoofing using Integrated Spectro-Temporal Graph Attention Networks)
• Graph-attention based architecture specialized for deepfake detection

---

Preprocessing Pipeline

A carefully designed signal processing chain improves robustness:

1. Resampling → 16 kHz mono
2. Butterworth Bandpass Filter → 70 Hz -- 8 kHz
3. Z-Score Normalization

This removes environmental artifacts and focuses the model on human vocal characteristics.

---

2-Way Test-Time Augmentation (TTA)

The system performs inference twice:

• Clean audio
• White-noise injected variant

Final score = average of both predictions

Prevents overfitting to digital silence
Improves robustness against adversarial audio

---

Tech Stack

• Python
• FastAPI
• PyTorch
• Librosa
• Hugging Face Transformers

---

Installation & Setup

Clone Repository

git clone https://github.com/yourusername/ai-voice-detector-api.git
cd ai-voice-detector-api

---

Install Dependencies

pip install -r requirements.txt

---

Set Environment Variable (API Key)

Linux / Mac:

export API_KEY=your_secret_key

Windows (PowerShell):

setx API_KEY "your_secret_key"

---

Run Server

python main.py

Server will start at:

http://0.0.0.0:7860

---

API Documentation

Endpoint

POST /api/voice-detection

---

Headers

x-api-key: <API_KEY>
Content-Type: application/json

---

Request Body

{
  "audioBase64": "<base64_encoded_audio>"
}

Supported formats:

• .wav
• .mp3

---

Response

{
  "status": "success",
  "classification": "AI_GENERATED",
  "confidenceScore": 0.9842
}

Classification values:

• HUMAN
• AI_GENERATED

---

Security

• API key authentication required
• No file uploads stored on server
• Stateless inference pipeline

---

Use Cases

• Voice fraud detection
• Biometric authentication security
• Deepfake forensics
• Call screening applications
• Research & experimentation

---

Performance Goals

• Low latency inference
• CPU compatible deployment
• Scalable containerization ready
• Minimal memory footprint

---

Project Structure

├── main.py
├── model/
├── utils/
├── requirements.txt
└── README.md

---

Attribution

This project builds upon the work of leading research organizations:

• Meta AI --- for the MMS-300M (Wav2Vec2) speech representation model
  https://ai.meta.com/

• Clova AI Research (Naver) --- for the AASIST architecture
  https://clova.ai/

If you use this project in research, please consider citing their work.

---

Disclaimer

This tool is intended for research and security purposes only.

Detection accuracy may vary depending on:

• Audio quality
• Compression artifacts
• Novel synthesis techniques

Always combine with additional verification methods for critical applications.

---

Contributing

Pull requests are welcome!

For major changes, please open an issue first to discuss what you would like to change.

---

License

MIT License

---

Author

Developed with care for AI Security & Fraud Prevention.