Abstract:

The increasing adoption of Generative Artificial Intelligence (GenAI) in education has opened new opportunities to optimise the production and transformation of teaching materials. Large Language Models (LLMs) can generate coherent, contextually grounded text adapted to multiple formats, making them promising tools for enhancing educational resources. However, most existing studies focus on content generation or summarisation, overlooking their potential to reconstruct and enrich degraded transcripts—an issue frequently observed in hybrid or remote teaching environments where audio quality, background noise or recognition errors significantly affect the accuracy of automated transcription.

This study examines the use of generative AI to process incomplete or noisy textual transcripts sourced from Google Meet recordings of a university extension course on the history and design of video games. The goal is twofold: to assess the ability of an LLM to reconstruct fragmented information while preserving semantic fidelity to the original lecture, and to quantify the level of correction, expansion and enrichment applied to each fragment through systematic analysis.

The methodology followed three stages. First, automatic transcripts containing syntactic, semantic and phonological errors were collected from several hybrid-format sessions. Second, the material was segmented into coherent thematic fragments (e.g., “8-bit consoles”, “video game crash”, “smartphones and casual games”), enabling fine-grained processing aligned with the structure of the lecture. Third, the model was instructed to act as an expert reconstruction system tasked with maintaining the lecturer’s tone, correcting major errors, completing unfinished sentences, adding natural discourse connectors, preserving technical references and enriching content when contextually required. Metrics such as original vs. resulting word counts, generated tokens, estimated linguistic correction and inferential intervention were recorded, and all reconstructed outputs were manually reviewed to ensure reliability.

Results using ChatGPT 5.1 show a clear positive impact of GenAI on transcript reconstruction. The final text expanded from 3,300 to approximately 4,900 words (+44%). The average correction rate was 39% per fragment, with peaks of 78% in highly degraded sections. More coherent fragments required minimal intervention (26–30%), while others with greater noise or thematic complexity demanded deeper reconstruction. The model also inferred omitted contextual information while maintaining the semantic integrity and narrative coherence of the original lecture. These improvements produced texts suitable for voice-over, course material generation and long-term archival.

This exploratory study demonstrates the potential of LLMs as linguistic and semantic enhancement tools in education. Their ability to transform imperfect transcripts into structured, pedagogically robust materials opens new avenues for managing audiovisual resources and automating content refinement processes. Future work includes integrating multimodal models, expanding the analysis to full-length recordings and comparing architectures according to transcript quality. Overall, generative AI emerges as a strategic asset for improving the accessibility, productivity and quality of university educational content.

Keywords:

Generative AI, Large Language Models (LLMs), Prompt Engineering, AI-Assisted Transcription, AI in Education.