🎙️ AI Voice Detector Engine (ResNet-18 Edition)

A professional-grade ResNet-Audio pipeline designed to detect synthetic speech across 5 languages with calibrated confidence.

Features • Architecture • Quick Start • API Reference • Model Details

---

🌟 Features

• Multilingual Support — Specialized for English, Hindi, Tamil, Telugu, and Malayalam.
• SE-ResNet Architecture — Squeeze-and-Excitation Attention mechanism dynamically weights Pitch vs. Spectral features for context-aware detection.
• Sliding Window Inference — Analyzes entire long-form dialogues by scanning 5-second overlapping chunks.
• Calibrated Confidence — Implements Temperature Scaling to ensure confidence scores are statistically honest.
• Real-World Robustness — Trained with Telephony Simulation (Bandpass/Gain), SpecAugment (Masking), and Channel Dropout (Zeroing Pitch/ZCR) to prevent overfitting.

---

🏗️ Architecture

┌─────────────────────────────────────────────────────────────────────────────┐
│                         AI VOICE DETECTION PIPELINE                         │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│   ┌──────────┐     ┌─────────────┐     ┌────────────────┐    ┌──────────┐   │
│   │ Audio    │────▶│ 11-Channel  │────▶│ Sliding Window │───▶│ Batch    │   │
│   │ (Base64) │     │ Feature Ext │     │ (5s Stride)    │    │ Inference│   │
│   └──────────┘     └─────────────┘     └────────────────┘    └──────────┘   │
│                                                                    │        │
│                                                                    ▼        │
│   ┌────────────┐     ┌────────────┐     ┌───────────┐     ┌─────────────┐   │
│   │ JSON       │◀────│ Temp Scale │◀────│ Max/Mean  │◀────│SE-ResNet-18 │   │
│   │ Response   │     │ Calibration│     │ Aggregator│     │ (Attention) │   │
│   └────────────┘     └────────────┘     └───────────┘     └─────────────┘   │
│                                                                             │
└─────────────────────────────────────────────────────────────────────────────┘

---

🚀 Quick Start

Prerequisites

• Python 3.10+
• Anaconda / Miniconda (Recommended)

Installation

cd ai-engine
pip install -r requirements.txt

Full Pipeline Workflow

# 1. Generate Multilingual Dataset
python scripts/generate_daily_ai_voices.py

# 2. Train the ResNet Model
python train.py

# 3. Calibrate Confidence Scores
python calibrate.py

# 4. Evaluate Performances
python evaluate.py

# 5. Start the Production API
python api.py

---

📡 API Reference

Detect Voice

POST /api/voice-detection
Content-Type: application/json
x-api-key: YOUR_API_KEY

{
  "language": "Tamil",
  "audioFormat": "mp3",
  "audioBase64": "<base64_encoded_audio>"
}

Response Format:

{
  "status": "success",
  "language": "Tamil",
  "classification": "AI_GENERATED",
  "confidenceScore": 0.9654,
  "explanation": "Very high confidence (96.54%) - Clear synthetic speech patterns detected. High spectral uniformity...",
  "meta": {
    "windows_analyzed": 4,
    "max_ai_prob": 0.9821,
    "avg_ai_prob": 0.8542
  }
}

---

🧠 Model Details

11-Channel Feature Extraction

The model doesn't just look at a spectrogram. It extracts 11 acoustic channels representing 113 unique features:

• Mel Spectrogram (64 bands)
• MFCCs + Deltas (26 channels)
• F0 Pitch Tracking (1 channel)
• Spectral Contrast (7 channels)
• Chroma STFT (12 channels)
• ZCR, Centroid, Bandwidth (3 channels)

Architecture: SE-VoiceResNet

A custom deep residual network enhanced with Squeeze-and-Excitation (SE) Blocks:

• Residual Blocks: 4 Layers of basic blocks for deep feature learning.
• Attention (SE): Adaptive average pooling blocks that learn to re-weight channels (e.g., ignoring Pitch if it's noisy) for every single inference.
• Dropout: 0.4 probability to prevent over-fitting.
• Focal Loss: Trained with Focal Loss to prioritize hard-to-classify examples over easy ones.

---

📊 Performance (Benchmark)

Metric	Score
Accuracy	97.08%
Recall (AI Detection)	100.00%
Precision (Human)	100.00%
ROC-AUC	0.9980

Note: Benchmarked on a balanced set of 565 samples across 5 languages.

---

📁 Project Structure

ai-engine/
├── 📄 api.py                  # Production FastAPI server
├── 📄 model.py                # VoiceResNet architecture
├── 📄 dataset.py              # Augmented Data Loader
├── 📄 train.py                # Training logic with early stopping
├── 📄 calibrate.py            # Calibration runner
├── 📄 calibration.py          # Temperature scaling implementation
├── 📄 inference.py            # Sliding window prediction engine
├── 📄 requirements.txt        # Dependencies
└── 📁 scripts/
    └── 📄 generate_daily_ai_voices.py  # Multilingual TTS generator

---

Built with ❤️ for robust synthetic voice detection.