At Data Pilot in Lahore I deployed NoiseVu, an edge-based sound classification system for crime prevention. The model runs on Raspberry Pi using Wave2Vec2 for robust acoustic event detection.
Improvements
- Accuracy: Raised detection accuracy from 70% to 92% through better training and calibration.
- Latency: 35% latency reduction by model quantization and on-device inference.
The pipeline captures audio on the Pi, runs the classifier locally, and triggers alerts or logs when target events are detected, without sending raw audio to the cloud.