This code is the culmination to my Capstone project at the University of Washington: Integrating Large Language Models in Verbal Autopsy Workflows. The project is built on the transformer platform from huggingface and utlises three different AI tasks, Speech-to-text, Zero-shot classification, and text generation.
- Speech-to-text using Whisper from OpenAI
- Text generation multiple options but most robustly tested using ChatGPT4 with the OpenAI api
- Zero-shot classification using Bart-large-mnli from Facebook
S-T-T is geared towards aiding VA administrators in the process of transciption and translation of interviews in near real time. Using the transcibed narrative, text generation and zero-shot classification approaches are used to try and classify a cause of death. By using the transformer platform, there is malleability in which models are chosen for classification allowing for new models to be adopted easily.