Wakeword detection on a humanoid robot

Context

Processing audio data with large artificial neural networks (ANNs) can be computationally intensive and power-consuming. To mitigate this, systems like Amazon Alexa employ a technique known as wakeword detection. A lightweight model constantly listens for a specific keyword ("Alexa"), activating the larger, more powerful models only when necessary.

In the JuBot project, developing a humanoid robot is being developed that requires a similar wakeword detection capability. This component is essential for ensuring that the robot responds only when directly addressed, optimizing both performance and power usage.

Tasks

• In this position, you will be responsible for implementing a wakeword detection component tailored to the JuBot humanoid robot. The system should be capable of identifying specific spoken commands, such as "Armar" to activate the robot or "Stop" to immediately halt its actions. A key part of the work involves ongoing improvement of the model to reduce false rejections and false acceptance.
• Development of a neural network, possibly based on BCResNet, for wakeword detection
• Integration into the ARMAR-X platform for interfacing with the robot
• Testing, maintenance, and finetuning

Prerequisites

• Strong Python foundations
• Strong deep learning foundations
• Basic PyTorch knowledge