Intelligent Acoustic Threat Detection Using Deep Learning and Classical Machine Learning
AuralGuard is an AI-powered sound classification system that detects suspicious environmental sounds — such as glass breaking, chainsaws, and door knocks — to support modern surveillance and public safety solutions. Built using the ESC-50 dataset, it compares traditional machine learning and deep learning models to find the most effective approach for acoustic-based threat detection.
- Detect and classify abnormal or hazardous environmental sounds
- Evaluate and compare multiple machine learning and deep learning models
- Enable future integration into smart city and real-time monitoring systems
- ESC-50: Environmental Sound Classification dataset by Karol J. Piczak
- Total: 2,000 labeled audio clips from 50 categories
- Classes used in this project:
glass_breakingchainsawdoor_wood_knock
- Dataset Source: ESC-50 GitHub
- Root Mean Square Energy (RMS)
- Zero Crossing Rate (ZCR)
- Spectral Centroid
- Spectral Bandwidth
- Mel-Spectrograms (for deep learning)
- ✅ Random Forest
- ✅ Decision Tree
- ✅ K-Nearest Neighbors (KNN)
- ✅ Convolutional Neural Network (CNN)
- ✅ Gated Recurrent Unit (GRU)
- ✅ Bidirectional GRU (BiGRU)
- Pitch Shifting (±2 semitones)
- Time Stretching (±10% speed)
- Spectrogram Normalization
| Model | Accuracy (%) |
|---|---|
| ✅ CNN (Deep Learning) | 98.33% |
| Bidirectional GRU | 96.67% |
| Random Forest | 87.50% |
| Decision Tree | 87.50% |
| SVM | 87.50% |
| K-Nearest Neighbors (KNN) | 68.75% |
✅ CNN achieved the highest accuracy, making it the best candidate for real-time deployment in acoustic surveillance systems.
- Confusion Matrices per Model
- Accuracy and Loss Curves (CNN, GRU, BiGRU)
- Horizontal Bar Chart: Model Accuracy Comparison
pip install -r requirements.txt