This research introduces a novel post-processing approach using fine-tuned Large Language Models to correct speaker diarization errors in conversation transcripts.

• Employs LLMs trained on the Fisher corpus to identify and fix speaker attribution mistakes
• Creates a generalizable solution that works across different conversation types
• Demonstrates significant improvements in diarization accuracy on holdout datasets
• Provides a practical enhancement layer for existing Automated Speech Recognition (ASR) systems

Linguistic Impact: This advancement is crucial for computational linguistics applications requiring accurate speaker identification, including conversation analysis, meeting transcription, and multi-speaker content processing.

LLM-based speaker diarization correction: A generalizable approach