Skip to main content

Deploying TensorFlow Lite on AudioMoth

Supervisor

Suitable for

MSc in Computer Science

Abstract

The use of convolution and deep neural networks on computationally constrained hardware has received lots of attention recently with the development of TensorFlow Lite which targets small ARM devices. This project will explore the use of this framework for audio detection tasks. The intended platform is AudioMoth -- a low-cost open-source smart acoustic sensor used for biodiversity and environmental sensing (https://www.openacousticdevices.info) — and a typical application would involve training a neural network to recognise the sound of a particular species of bird such that only recordings of that species are captured by the device. This project will develop a complete toolchain to allow a model to be trained and tested on a desktop machine using library recordings, and then ported to run on the AudioMoth device (constrained by its limited RAM and flash storage).