A $10 Arduino AI Chip Can Now Identify Mosquitoes by Buzz

Researchers including Kiran Trivedi from the University of Wollongong presented a tiny $10 device at the UN AI for Good summit in Geneva that identifies disease-carrying mosquitoes in real time with 88.3% accuracy purely from the sound of their wingbeats.

Mosquitoes might sound like uniform nightmare fuel to a sleeping human, but species like Anopheles and Aedes actually flap their wings at distinct acoustic frequencies and harmonic signatures. The team converted those raw audio signals into spectrogram images, effectively turning an audio classification task into a computer vision problem that neural networks love.

Instead of relying on cloud data centers, the system runs completely offline on microcontrollers that have mere kilobytes of RAM rather than gigabytes of VRAM. The full model was trained on GPU clusters before undergoing FP32 to INT8 quantization and pruning, stripping away useless weights until the entire neural network shrank to roughly 100 to 200 kilobytes compiled via TensorFlow Lite for Microcontrollers directly onto the hardware chip.

Deploying this outside controlled labs requires MEMS directional microphones to ignore tropical rain storms, combined with wake-on-sound power modes to stretch battery life in deep wilderness. The ultimate plan involves deploying distributed IoT sensor networks that transmit tiny telemetry packets to map disease vectors before outbreaks start.

Cloud AI companies keep begging for nuclear power plants to answer basic search queries, while ten-dollar microcontrollers silently save lives on a single coin battery in the jungle.

Source: University of Wollongong