Intent Recognition in Speech-to-Text Processing in the Context of Natural Interaction with cognitive assistive Systems

This study develops an efficient intent detection pipeline for patient-robot interaction in healthcare, optimized for resource-constrained devices like the Jetson AGX Orin. Using a custom dataset of simulated nursing home interactions with text and speech. The research evaluates 10 ASR systems, 20 LLMs, and 4 LALMs. Integrated ASR + LLM and end-to-end LALM pipelines were tested for speech-to-text intention recognition. The analysis indicates that the 2-step integrated ASR + LLM pipelines outperform end-to-end LALM technologies for intent detection in speech-to-text processing within the defined use case, achieving higher accuracy while maintaining more efficient runtime and resource consumption.