NUTRI-CORRECT: KNOWLEDGE GRAPH-AIDED GENERATIVE ERROR CORRECTION FOR PRECISION NUTRITION SPEECH RECOGNITION

Abstract

Applications that make use of Automatic Speech Recognition (ASR) technology are now widely used within the health care and nutrition sectors; however, existing ASR models have been known to frequently make mistakes when recognizing specialized ter minology and phrases related to health and nutrition. Such problems may pose issues for health care applications such as dietary analysis, telemedicine services, and nutri tion intelligence assistants. Therefore, this thesis will explore ”Nutri-Correct: Knowledge Graph-Aided Generative Error Correction for Precision Nutrition Speech Recognition.” The proposed approach combines a pre-trained ASR model with a nutrition-specific knowledge graph and a generative correction module based on transformers. First, speech is transcribed with the help of an ASR baseline engine, followed by post-processing with contextual information about nutrition and generative language models to fix errors re lated to semantics or phonetic similarity in the context. This architecture is aimed at overcoming the challenges related to pronunciation dif ferences, ambient noise, and ambiguity inherent to the medical domain. The experiments showed that the framework provides significantly better transcription quality than tra ditional ASR models in terms of the WER, Precision, Recall, F1-score, and Semantic Accuracy metrics. The Nutri-Correct system can be used to develop applications like virtual nutritionists, nutritional monitoring, telemedicine services, and AI-supported clinical decision-making systems. This paper showed how hybrid neuro-symbolic approaches could improve spe cialized speech transcription tasks in the precision nutrition domain.