Disfluency Detection and Removal in Speech Transcriptions via Large Language Models
Pedro L. S. de Lima, Cláudio E. C. Campelo · 2024
The field of Automatic Speech Recognition (ASR) has significantly expanded within the technological landscape due to its extensive use in sectors such as education, healthcare, and customer service. Many modern applications depend on analyzing spoken content through Speech-to-Text (STT) conversion models. However, transcriptions produced by these systems often contain undesirable elements, such as word repetitions and the prolongation of certain sounds, known as disfluencies or linguistic crutches. These elements can negatively affect the quality of automatic content analysis by Natural Language Processing (NLP) models, including those for named entity recognition, emotion detection, or sentiment analysis. Therefore, this study aims to evaluate the feasibility of identifying and eliminating linguistic disfluencies using Large Language Models (LLMs), such as GPT-4, LLaMA, Claude, and Gemini, through Prompt Engineering techniques. The approach was tested using a corpus of debate transcriptions with manually annotated disfluency occurrences, yielding promising results.