Details

Automatic Speech Recognition Adaptation for Pharma IoT Devices

Year: 2026

Term: Winter

Student Name: Hakan Khan

Supervisor: Omair Shafiq

Abstract: Over half of patients taking medications have non-adherence to their medications, which causes a third of hospital visits. A solution to the problem of weak pharmaceutical adhesion will ease the patient’s burdens twofold, as within Canada, our public healthcare workforce is stretched thin, and having a dedicated nurse, or nursing home is expensive for those who require medication. Additionally, interacting via voice commands is the most natural way for humans to communicate. Therefore, it is the best solution to ensure the widest coverage when designing an IOT device for adoption by the widest range of users. The goal of this project was to train an Automatic Speech Recognition (ASR) system on a vocabulary spanning all the pharmaceutical terminology in the FDA database to bring accuracy to a comparable word error rate to the state-of-the-art ASRs on the market. Since these rare pharmaceutical words do not occur in common data sets used for training the models initially, these ASRs have traditionally struggled to interpret these. We have employed fine-tuning and evaluation to fulfill this goal.