ANR LIGHT-SWIFT
Embedded acoustic intelligence for low-power connected sensing
This project explores how far we can push real-time acoustic understanding on resource-constrained embedded platforms. The device captures audio continuously, extracts compact features on-board, runs quantized classification locally, and pushes actionable results through a low-latency wireless path.
Visit the ANR project pageSystem Architecture
Click Acoustic Event to start 路 Event Detection toggles the inference pipeline
Processing Pipeline
Performance Benchmarks
Model Size vs Accuracy
Energy per Inference
5x energy savings with NPU + cache optimization
Latency Breakdown (Total: ~120ms)
Key Achievements
End-to-end latency from sound capture to BLE alert
Classification accuracy with INT8 quantized model
Energy savings with NPU vs CPU inference
Project Timeline
December 2023
Project Start
Official launch of the ANR-LIGHT-SWIFT collaborative research project between WAVELY, Inria, and NII (Tokyo).
March 18-19, 2024
Kickoff Meeting
First consortium meeting gathering all partners to define research objectives, work packages, and initial milestones.
April 7-9, 2025
Face-to-Face Meeting - NII Tokyo
On-site collaboration in Tokyo to present embedded inference results, dataset alignment, and shared experimentation.
November 13, 2025
IEEE Publication
Research paper published on IEEE Xplore presenting the embedded acoustic classification pipeline and its energy-focused design.
View on IEEE Xplore