Research on Real-time Multilingual Transcription and Minutes Generation for Video Conferences Based on Large Language Models

This paper presents an innovative approach to real-time multilingual transcription and minutes generation for video conferences using Large Language Models (LLMs). The proposed system integrates advanced speech recognition techniques with sophisticated natural language processing capabilities to address the challenges of multilingual communication in virtual meetings. The implementation incorporates a novel hierarchical architecture combining transformer-based models for speech recognition and rhetorical structure modeling for automated minutes generation. The system achieves significant performance improvements with an average Word Error Rate of 4.2% across supported languages and ROUGE-L scores of 0.825 for minutes generation. Through the implementation of adaptive resource allocation and selective forwarding techniques, the system demonstrates a 35% reduction in bandwidth consumption while maintaining processing latency under 150 milliseconds. The paper introduces a comprehensive evaluation framework incorporating both automated metrics and human assessment, demonstrating robust performance across various operational conditions. Experimental results show improvements in transcription accuracy by 28% and resource utilization efficiency by 25% compared to baseline systems. The system supports simultaneous processing of five major languages while maintaining consistent performance levels across different meeting scenarios. The research contributes to the advancement of multilingual video conferencing technology by providing a scalable and efficient solution for real-time communication and documentation needs.