AeroShield is a hybrid AI + DSP noise cancellation system designed for real-time speech enhancement in defence, aerospace, and high-noise industrial environments. It combines classical adaptive filtering (NLMS) with a capacity-optimized RNNoise deep learning model, deployed on low-power embedded hardware (NXP MCX N236, ARM Cortex-M33) — delivering real-time noise suppression at a fraction of the cost of GPU-class edge AI platforms.
Defence and mission-critical communication systems are severely affected by diverse acoustic disturbances — gunshots, artillery fire, helicopter rotor noise, armored vehicle sound, and emergency sirens. Traditional signal processing techniques assume stationary noise and struggle under highly dynamic, non-linear conditions, often introducing artifacts or speech distortion.
AeroShield addresses this with a hybrid AI/ML-driven noise suppression pipeline combined with adaptive filtering, deployed on cost-effective, low-power embedded hardware, targeting real-time performance suitable for defence, aerospace, and high-noise industrial communication.
- 🌐 Live Website: https://frdm-five.vercel.app/
AeroShield processes incoming audio through three stages:
- Signal Conditioning (DSP Pre-Filter) — dual-microphone (primary + reference) NLMS adaptive filtering removes correlated background noise, followed by DC-blocking and high-pass filtering.
- Neural Noise Suppression — a quantized, capacity-optimized RNNoise model performs real-time spectral gain estimation to suppress complex, non-stationary, and impulsive noise.
- Output Refinement (DSP Post-Filter) — automatic gain control, and soft peak limiting restore harmonic clarity and ensure consistent, distortion-free output.
The system is deployed end-to-end on an NXP MCX N236 (ARM Cortex-M33) microcontroller, with the neural inference path optimized via ARM assembly for single-core, real-time execution.
├── software/ # Website — results explorer, live hardware telemetry, documentation, interactive home page
├── ml/ # Dataset generation, model training, quantization, and evaluation
└── hardware/ # Embedded DSP pre/post-filters, ARM-assembly-optimized RNNoise inference, linker configuration
- NXP MCX N236 (ARM Cortex-M33)
- Primary + reference microphones
- External DAC (audio output)
- Supporting peripherals as detailed in
hardware/README.md
- Want to see it in action? → Visit the live website
- Want to train or evaluate the model yourself? → See
ml/README.md - Want to build and flash the hardware? → See
hardware/README.md - Want to run or modify the website? → See
software/README.md