EPIC: Error Pattern Informed Correction for Classroom ASR with Limited Labeled Data

Abstract

Automatic speech recognition (ASR) systems have a wide range of applications in classroom analysis. However, due to the unique structure of classroom dialogue, existing ASR systems often struggle to accurately recognize and organize spoken utterances, creating significant challenges for downstream tasks in educational dialogue analysis. To address this issue, we propose EPIC, a post-processing framework for classroom ASR error correction. We begin by extracting error patterns to gain a deeper understanding of the distribution of ASR errors. Next, we utilize large language models (LLMs) to reconstruct contextual information based on these error patterns, offering a viable solution for error correction with limited labeled data. Finally, after fine-tuning an error correction model, we implement a candidate selection process to identify the most appropriate hypothesis for each context. Extensive experiments with our proposed method demonstrate substantial improvements in word error rate (WER) and overall robustness in ASR error correction, enabling more reliable analysis of educational dialogues and offering deeper insights for educational research.