Course Stage Recognition for Online Course Recordings Using Spoken Language Understanding

EasyChair Preprint 15289

Abstract

This study investigates models for course stage recognition, a novel task in Spoken Language Understanding (SLU) aimed at segmenting classroom recordings into distinct instructional phases. Two approaches are evaluated: an end-to-end SLU model based on the WavLM base+ speech encoder, and a multistage SLU method integrating Whisper for Automatic Speech Recognition and ChatGPT 4o for Natural Language Understanding. The study compares the performance of these models to explore stage recognition without relying on intermediate text representations. Results indicate that the multistage approach excels in fine-grained classification across five stages—Opening, Lecture, Break, Conclusion, and Others—but is outperformed by the end-to-end model in distinguishing the Lecture stage. The findings suggest that a speech-language model capable of performing in-context learning directly on speech data could further enhance the accuracy of course stage recognition.

Keyphrases: Course Stage Recognition, Large Language Model, Speech Model, Spoken Language Understanding

@booklet{EasyChair:15289,
  author    = {Yi-Ting Yuan and Ke-Ching Hong and Yu-Hsiang Wang},
  title     = {Course Stage Recognition for Online Course Recordings Using Spoken Language Understanding},
  howpublished = {EasyChair Preprint 15289},
  year      = {EasyChair, 2024}}