Projects

Bird Audio Classification on Edge Devices - TinyML Pipeline

TinyML pipeline for low-power bioacoustic classification, connected to research accepted at IEEE ICASSP 2026.

Highlights

• Developed an end-to-end TinyML pipeline trained on 150k+ audio samples across 70 species.
• Designed lightweight CNN-GRU architectures optimized for microcontroller deployment, reaching up to 90.8% accuracy on distinctive species.
• Deployed real-time inference on AudioMoth firmware using TensorFlow Lite Micro and CMSIS-NN, achieving about 16x lower energy consumption than baseline systems.