Abstract:
Voice biometrics use the unique vocal attributes of a user to secure physical, virtual,
and online assets. Current voice biometric systems use cloud data centers to pro
cess voice samples received from computers or mobile devices[1]. However, in the
near future, it will be possible to use programmable voice biometric chips that can
leverage pre-trained machine learning models to operate autonomously. Such chips
would make it possible to deploy voice biometric systems more widely to secure
physical assets without the need for internet connections[2].
This project introduces a voice biometric system tailored for home or office security
applications. The system’s core functionality involves detecting and verifying user
voices by processing multiple instances of wake-up words presented as 4-digit pass
codes. Designed to authenticate various users, the project progressed through stages
that encompassed researching and evaluating diverse voice extraction methods. The
optimal method was selected for the system. Subsequently, speaker recognition
and spoken digit classification models were implemented and assessed. The study
achieved significant milestones by rigorously testing, simulating, accelerating, and
deploying the models onto hardware, yielding promising results and advancements.
