Implementation of the Bidirectional Encoder Representations From Transformer Algorithm in Speech To Text for Meeting Minutes

Abstract

Meeting transcription is a crucial process for organizations, yet it often consumes significant time and resources due to the manual effort involved in recording, understanding, and documenting discussions accurately. In the digital era, advancements in speech processing and natural language understanding provide an opportunity to automate this process. This research focuses on the implementation of the Bidirectional Encoder Representations from Transformers (BERT) algorithm in a Speech-to-Text (STT) system to enhance the accuracy and efficiency of meeting transcriptions. The study integrates BERT, a deep learning-based model capable of comprehending bidirectional contextual information, into the transcription pipeline to improve handling of complex conversational contexts. The research follows a systematic methodology, starting from data preprocessing, model training, and evaluation to assess its performance. Results show that the proposed system achieves high transcription accuracy, demonstrating significant potential for real-world applications in organizational environments. This research also highlights the importance of advanced NLP technologies, such as BERT, in overcoming challenges of transcription in multilingual and noisy environments. The developed system offers practical benefits in terms of reducing manual effort and improving access to meeting documentation, making it a valuable tool for productivity enhancement.