Speech Keyword Spotting for Unmanned Ground Vehicles

This project aims to design a keyword spotting system (KWS) using microphone arrays installed on an unmanned ground vehicle (UGV). The project consists in 1) installing microphone arrays on the partner’s UGV, recording the ego-noise generated by the wheels and engine, and measuring the acoustic transfer function for each possible direction of arrival; 2) creating a speech commands dataset using a custom-made website for crowd sourced keyword collection; 3) performing data augmentation using the recorded noise and acoustic transfer functions; 4) designing and training a convolutional recurrent neural network using the augmented data to detect and localize keywords; 5) deploying the neural networks on embedded hardware; 6) validating the system on the UGV in real-life conditions. This project will provide the industrial partner with a unique voice interface to assist users in adversarial acoustic environments and will push forward the field of joint keyword spotting and localization using deep learning.

Faculty Supervisor:

François Grondin

Student:

Partner:

Rheinmetall Canada Inc

Discipline:

Engineering

Sector:

Manufacturing

University:

Université de Sherbrooke

Program: