INSPECTA 1.0: IMPLEMENTATION CHALLENGES FOR ON-DEVICE
SPEECH AND VISION TASKS

Artificial intelligence (AI) and Machine Learning (ML) has become ubiquitous in our day-to-day lives, powering common conveniences such as smart home controls and entertainment suggestions. However, AI has not yet seen wide deployment for safeguards related tasks. Inspecta, the International Nuclear Safeguards Personal Examination and Containment Tracking Assistant, is being developed to reduce the burden of common safeguards tasks encountered during inspections. A key focus of Inspecta is to leverage existing AI/ML techniques to improve the inspection experience rather than developing new algorithms. The initial 1.0 version of the Inspecta Android application has the capability to perform on-device machine learning for several tasks including speech recognition, speech synthesis, and optical character recognition (OCR). This paper documents challenges and solutions for these tasks and how they are applied to improve seal verification.