Designing an LLM-based Assistant for Speech Therapists: the Case of e-SpeechT

The Large Language Models (LLMs) are increasingly being employed across many fields, such as healthcare, finance, and automotive, revolutionizing how humans carry out activities. In this research, the integration of an LLM-based assistant in e-SpeechT, a web application that supports speech therapists in conducting, administering, and monitoring therapies, is investigated. It is intended to integrate an LLM component that allows the therapist to be supported in the generation of exercises to administer during the therapy, establishing a symbiotic relationship between the two parties. This approach reflects the core ideas of End-User Development (EUD) by enabling therapists to directly influence, customize and manage system behaviour without relying on advanced programming knowledge. The system enables the therapist to share the created exercise, reinforcing the concept of Cultures of Participation (CoP). Using Prompt Engineering principles and techniques, many prompts have been defined and underwent a multi-stage evaluation pipeline to assess the quality of the generated outputs. The obtained results show that a well-structured and clear prompt can generate consistent and reliable outputs that align with the prompt instructions.