Phone Call Deepfake Detector

Real-time AI voice detection optimized for phone call quality audio. Unlike most detectors that fail on telephony audio, this one is built specifically for the compression, noise, and limited bandwidth of actual phone calls.

Why This Exists

Most voice cloning scams happen over phone calls, but most AI detectors are trained on clean studio audio. Phone calls have:

• 8kHz sample rate (vs 44kHz normal audio)
• Narrow frequency band (300Hz - 3.4kHz)
• Heavy compression (GSM, AMR codecs)
• Noise, echo, packet loss

This detector is trained on phone-degraded audio to catch deepfakes where they actually happen.

Features

• Phone-optimized model - Trained on 8kHz, codec-compressed audio
• Real-time detection - Works during live phone calls
• On-device processing - No data leaves your phone
• Low latency - Results in <200ms
• 4 model architectures - From lightweight mobile to full attention-based

Architecture

Phone Call (8kHz)
      |
      v
+------------------+
| Noise Reduction  |
+--------+---------+
         |
         v
+------------------+
| Mel Spectrogram  |  64 bands, phone frequency range
+--------+---------+
         |
         v
+------------------+
| Lightweight CNN  |  Optimized for mobile
+--------+---------+
         |
         v
   REAL / FAKE + confidence %

Phone Degradation Pipeline

Training data is created by degrading clean audio to simulate phone calls:

clean_audio
    -> Resample to 8kHz
    -> Apply GSM codec simulation
    -> Bandpass filter (300Hz - 3.4kHz)
    -> Add background noise (SNR 10-30dB)
    -> Simulate packet loss (0-5%)
    -> Add echo/reverb

Quick Start

1. Setup Environment

cd deepfake-phone-detector
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

2. Prepare Dataset

# Setup directories and get instructions
python scripts/download_dataset.py

# If using ASVspoof dataset:
python scripts/prepare_asvspoof.py --zip data/downloads/LA.zip

# Apply phone degradation
python scripts/degrade_to_phone.py --input data/raw --output data/phone

3. Train Model

python model/train.py --data data/phone --epochs 50 --model mobile

4. Test Detection

python scripts/test_detection.py audio.wav --checkpoint checkpoints/best.pt --model mobile

5. Export for Mobile

python model/export_tflite.py --checkpoint checkpoints/best.pt --output android/app/src/main/assets/model.tflite

6. Build Android App

cd android
./gradlew assembleDebug
# APK: android/app/build/outputs/apk/debug/app-debug.apk

Project Structure

deepfake-phone-detector/
├── model/
│   ├── phone_augment.py    # Phone degradation pipeline (8kHz, GSM, noise)
│   ├── dataset.py          # PyTorch dataset for phone-quality audio
│   ├── model.py            # 4 CNN architectures (PhoneCNN, Mobile, Temporal, Attention)
│   ├── train.py            # Training loop with validation
│   └── export_tflite.py    # ONNX -> TFLite conversion
├── android/
│   ├── app/src/main/
│   │   ├── java/.../       # Kotlin: MainActivity, AudioService, Classifier
│   │   └── res/            # UI layouts and resources
│   └── build.gradle
├── scripts/
│   ├── download_dataset.py   # Dataset setup instructions
│   ├── prepare_asvspoof.py   # ASVspoof dataset extraction
│   ├── degrade_to_phone.py   # Batch phone degradation
│   ├── generate_bark_samples.py  # Generate fake samples with Bark TTS
│   └── test_detection.py     # Inference script
├── data/                     # Dataset (gitignored)
│   ├── raw/real/            # Original real voice samples
│   ├── raw/fake/            # Original AI-generated samples
│   ├── phone/real/          # Phone-degraded real samples
│   └── phone/fake/          # Phone-degraded fake samples
└── requirements.txt

Model Architectures

Model	Parameters	Inference	Use Case
phone_cnn	30K	15ms	Ultra-lightweight baseline
temporal	85K	25ms	Better temporal artifacts
mobile	17K	10ms	Android deployment
attention	443K	40ms	Highest accuracy

Dataset Sources

Real Voice Samples

• ASVspoof 2019/2021 - bonafide samples
• LibriSpeech - clean speech
• VoxCeleb

Fake (AI-Generated) Samples

• ASVspoof 2019/2021 - 17 spoofing systems
• Generate with Bark TTS (included script)
• Generate with Coqui TTS
• ElevenLabs API

Tech Stack

• ML: PyTorch 2.x, torchaudio
• Audio Processing: librosa, scipy, audiomentations
• Mobile: TensorFlow Lite, Kotlin
• Phone Simulation: Custom GSM codec simulation, bandpass filters

Performance

Dataset	Accuracy	Precision	Recall	F1
ASVspoof (clean)	~95%	~94%	~96%	~95%
ASVspoof (phone)	~92%	~90%	~93%	~91%
In-the-wild	~88%	~85%	~90%	~87%

License

MIT