With the emergence of powerful speaker recognition systems such as Alexa and Siri, people are becoming fond of using voice activation instead of physical input. Together with the rapid growth of the Internet of Things (IoT), an increasing number of embedded devices are being deployed in various settings that could benet from voice activation.
However, embedded IoT devices are often not as powerful as Alexa or Siri, and only have a fraction of their memory and processing power. Our goal is to let resource-constrained embedded IoT devices (i.e., deeply embedded devices with limited memory and computational resources) perform speaker recognition autonomously: this could be used to detect people's presence in a room (e.g., to easily clock in and out of work), or to eliminate the need for bulky keyboards or pin code inputs. Instead, only a small and cheap microphone is needed. Towards this goal, we have shrunk a state of the art model to fit the resource constraints of low-power IoT devices while retaining on-par accuracy. We are now interested on its deployment, on improving its robustness and usability, and on benchmarking its performance using different platforms, such as the OpenEarable 2 (a fully open-source AI platform for ear-based sensing applications) and the NPU-enabled Grove Vision AI Module V2.
Download as PDF
Student Target Groups:
- Students of ICE/Telematics;
- Students of Computer Science;
- Students of Electrical Engineering.
Thesis Type:
- Master Project / Master Thesis
Goal and Tasks:
Within this context, students can explore several directions and perform different tasks, such as:
- Understand how state-of-the-art speaker recognition systems work, and how they can be shrunk to fit on embedded devices;
- Explore different enhancements to improve robustness and usability, for instance, training-data augmentation techniques, comprehensive benchmarking under varied acoustic conditions, or alternative model-training strategies;
- Leverage a second microphone aimed toward the ear to separate external noise from speech and perform active noise-cancellation, and other sensors to conrm that the wearer is speaking;
- Develop a robust speaker-recognition system prototype running on a next-generation IoT device equipped with an on-chip Neural Processing Unit (e.g., an integrated Arm Ethos-U55).
Recommended Prior Knowledge:
- Good knowledge of machine learning;
- Good skills in Python and C programming;
- Experience with embedded microcontrollers.
Start:
Contact: