ANR LIGHT-SWIFT
Embedded acoustic intelligence for low-power connected sensing
This project explores how far we can reduce consumption of machine learning models (especially for real-time acoustic classification) on resource-constrained embedded platforms. The device tested captures audio continuously, extracts compact features, and runs a quantized neural network to classify environmental sounds with low latency and minimal energy consumption.
Visit the ANR project pageSystem Architecture
Activate the diagram to follow the signal from sound capture to wireless alerting.
Click Acoustic Event to start
Processing Pipeline
Performance Benchmarks
In-distribution metrics for model compression techniques
QAT-INT8 is the best trade-off between model size and classification performance. The model is 9.7x smaller than the baseline model while simultaneously improving the F1-score to 82.5%. Thanks to quantization-aware training (reducing overfitting during training), the model can be deployed on ultra-low-power hardware.
Energy per Inference (compressed model QAT-INT8)
On-device measurement for QAT-INT8 model deployment, the hardware-accelerated path cuts inference energy by more than 7x.
Latency Breakdown (Total: ~290 ms)
Key Achievements
Batch formation and on-device processing, before BLE notification.
In-distribution F1-score of the quantized model (QAT-INT8) on in-situ dataset (83.20% coverage for confidence index >60%).
Reduction versus a CPU-only inference execution path.
Project Timeline
December 2023
Project Start
Official launch of the ANR LIGHT-SWIFT collaborative research program.
March 18-19, 2024
Kickoff Meeting
The consortium aligned work packages, research priorities, and initial technical milestones.
April 7-9, 2025
Face-to-Face Meeting in Tokyo
Results on embedded inference, acoustic data alignment, and experimentation were shared with partners.
November 13, 2025
IEEE Publication
Publication of a peer-reviewed paper presenting the state-of-the-art in machine learning model compression techniques for energy-efficient deployment.
View on website IEEE XploreJune 22-24, 2026
Face-to-Face Meeting in Lannion
Annual consortium meeting to review progress and plan next steps. Planned the next proof-of-concept (POC) for September 2026 in Nagoya (Tokyo).