🛡️ VoiceGuard
Real-Time AI Detection & Prevention of Voice Cloning Impersonation Attacks
SIH 2026 · Problem Statement SIH26104
A sub-200ms edge detection engine that analyzes raw audio waveforms to detect and terminate synthesized deepfake calls — purpose-built for India's multilingual telecom ecosystem.
📌 Table of Contents
- 🚨 The Problem
- 💡 Our Solution
- 🧠 Model Architecture
- ⚡ Performance Benchmarks
- 🚀 Quick Start
- 🔌 API Reference
- 🖥️ Live Demo
- 🏗️ System Design
- 🛡️ Mitigation Workflow
- 🛠️ Tech Stack
- 👨👩👧👦 Team
- 📋 Disclaimer
🚨 The Problem
Voice cloning technology has made it trivial to impersonate anyone. A 10-second sample of a person's voice is enough to train a model that can speak anything in their voice.
In India, this has fueled a rising wave of financial fraud, misinformation, and identity theft carried out over ordinary phone calls:
- 📞 "Mom, I've been kidnapped" scams using a cloned child's voice
- 🏦 OTP / bank executive frauds with cloned official voices
- 🗳️ Political misinformation spread via fabricated audio of public figures
- 🇮🇳 Multilingual attacks exploiting gaps in accent- and language-agnostic detection
Existing defenses are slow, GPU-hungry, and fail on Indian languages and accents.
💡 Our Solution
VoiceGuard is a real-time, streaming deepfake detection system that:
| Capability | Detail |
|---|---|
| ⚡ Real-time | Sliding 250ms windows, verdicts in <185ms |
| 🗣️ Raw waveform analysis | 1D Sinc-Conv learns acoustic features directly — no mel-spectrogram bottleneck |
| 🧊 Edge-optimized | INT8 dynamic quantization + Conv-BN fusion = runs on 2 vCPU cores, 2.2MB effective footprint |
| 🇮🇳 India-first | Trained on samples across 12 Indic languages, zero accent bias |
| 🚨 Autonomous kill-switch | Auto-terminates calls at 75% synthetic confidence over 2 consecutive windows |
| 🔐 Privacy-preserving | No raw audio stored; SHA-256 chunk hashing; on-premise inference |
Detection Classes
| Class | Meaning |
|---|---|
gt |
Ground truth — bona-fide human speech ✅ |
synthetic_indic |
TTS-synthesized Indian-language speech 🚨 |
diffwave |
DiffWave neural vocoder 🚨 |
melgan |
MelGAN neural vocoder 🚨 |
wavenet |
WaveNet neural vocoder 🚨 |
🧠 Model Architecture
VoiceGuardRawNet — a compact quantized 1D RawNet designed for CPU-only, low-latency edge deployment:
Input: Raw 16 kHz waveform (1.5s window = 24,000 samples)
│
├── SincConv (64 learnable sinc band-pass filters, k=129)
├── ResidualBlock (64 → 64)
├── ResidualBlock (64 → 128, stride 2)
├── ResidualBlock (128 → 128)
├── MaxPool1d (k=10, stride 10)
├── BiGRU (2 layers, hidden 256) ──► INT8 quantized
└── Linear (512 → 5 classes)
Key optimizations applied:
- 🔹 Learnable sinc band-pass filters replace hand-crafted features
- 🔹 Conv-BN fusion for reduced inference latency
- 🔹 INT8 dynamic quantization on GRU + Linear layers
- 🔹 Sinc kernel pre-caching eliminates runtime
torch.sinc()calls - 🔹 ~206K parameters → extremely lightweight
⚡ Performance Benchmarks
Model footprint and latency measured on 2 vCPU threads (no GPU):
| Metric | Baseline | Fused | Fused + INT8 | Budget |
|---|---|---|---|---|
| P95 Latency | 126 ms | 160 ms | 138 ms | ✅ <185 ms |
| Model Size | 1.0 MB | — | 1.02 MB | ✅ <2.2 MB |
| Mean Latency | ~104 ms | ~128 ms | ~110 ms | ✅ |
📊 Both latency and footprint stay comfortably within SIH deployment budgets — no GPU required.
🚀 Quick Start
1. Clone the repository
git clone https://huggingface.co/indrajit4533/voiceguard
cd voiceguard
2. Install dependencies
pip install -r requirements.txt
3. Run the server
python app.py
4. Open the dashboard
http://localhost:8000
🎯 The live dashboard includes an audio-upload interface where you can test real files and see class probabilities, threat scores, and inference latency instantly.
🔌 API Reference
GET / — Landing dashboard
Serves the interactive VoiceGuard web UI — upload audio, see live detection results.
GET /api/health
{
"status": "healthy",
"model_loaded": true,
"model_type": "VoiceGuardRawNet INT8",
"classes": ["synthetic_indic", "diffwave", "melgan", "gt", "wavenet"]
}
POST /api/analyze
Upload audio (multipart/form-data, field: file):
curl -X POST http://localhost:8000/api/analyze \
-F "file=@sample.wav"
Response:
{
"threat_score": 96.4,
"authenticity_score": 3.6,
"top_class": "melgan",
"action": "TERMINATE CALL",
"is_synthetic": true,
"probabilities": {
"synthetic_indic": 1.2,
"diffwave": 0.8,
"melgan": 96.4,
"gt": 3.6,
"wavenet": 0.0
},
"inference_ms": 132.5
}
🖥️ Live Demo
Try it with:
- 🎙️ Your own voice recording → expect ALLOW CALL
- 🔊 Any TTS / AI voice clip → expect TERMINATE CALL
- 📁 Real-world deepfake audio datasets (ASVspoof, MLAAD, etc.)
🏗️ System Design
┌────────────┐ ┌──────────────┐ ┌──────────────────┐ ┌────────────┐
│ Inbound │ ──► │ Streaming │ ──► │ VoiceGuard RawNet │ ──► │ Verdict │
│ VoIP Call │ │ Ring Buffer │ │ INT8 Inference │ │ Pass/Drop │
│ │ │ (250ms win) │ │ 2 vCPU, <185ms │ │ │
└────────────┘ └──────────────┘ └──────────────────┘ └────────────┘
│ │ │ │
│ │ │ ▼
│ │ │ ┌──────────────┐
│ │ │ │ Kill-Switch │
│ │ │ │ ≥75% × 2w │
│ └───────────┐ │ └──────────────┘
│ │ │ │
▼ ▼ ▼ ▼
┌────────────┐ ┌────────────┐ ┌────────────┐ ┌─────────────┐
│ SHA-256 │ │ Threat │ │ Audit Log │ │ Call │
│ Chunk Hash │ │ Scorer │ │ (Forensic) │ │ Termination │
└────────────┘ └────────────┘ └────────────┘ └─────────────┘
Streaming pipeline
- Ingest — raw VoIP stream partitioned into 250ms chunks (~4,000 samples @16kHz)
- Buffer — sliding window accumulates 1.5s (24,000 samples) for context
- Detect — INT8 model scores each window; threat =
1 − P(bona_fide) - Mitigate — sustained ≥75% confidence across 2 consecutive windows triggers automatic call termination
🛡️ Mitigation Workflow
High Threat Detected (≥75%)
│
▼
┌──────────────────┐
│ Kill-Switch Arm │ ── 1st consecutive window above threshold
└──────────────────┘
│
▼
┌──────────────────┐
│ Second Window? │
└──────────────────┘
│ No │ Yes
▼ ▼
Re-analyze ┌──────────────────┐
│ TERMINATE CALL │
│ + Forensic Audit │
│ + Alert Dispatch │
└──────────────────┘
🛠️ Tech Stack
| Layer | Technology |
|---|---|
| Backend | FastAPI + Uvicorn |
| Model | PyTorch (1D RawNet, INT8 dynamic quantization) |
| Audio | TorchAudio |
| Web UI | HTML / Tailwind CSS / vanilla JS |
| Serialization | Safetensors (pickle-free, secure) |
| Deployment | Hugging Face / Docker / local edge servers |
👨👩👧👦 Team
VoiceGuard is built by Team's for Smart India Hackathon 2026 (SIH26104):
| 👩💻 Member | Role |
|---|---|
| Indrajit Yadav | Founder & Lead Developer · Model Design, Backend, Quantization & Deployment |
| Swetha S | Team Member · Model support & Evaluation |
| Harinandan P S | Team Member · Pipeline & Integration Support |
| Dharshini M | Team Member · UI/UX & Design |
| Sudarsan | Team Member · Testing & Deployment Support |
| Divya | Team Member · Research & Documentation |
📋 Disclaimer
This project is provided for research and ethical security applications under Smart India Hackathon 2026. VoiceGuard is designed to detect and protect against voice-cloning fraud — it does not generate synthetic speech. Use responsibly.
Built with ❤️ for India's telecom ecosystem
Smart India Hackathon 2026 · Problem Statement SIH26104