Toward a More Accurate Speech-to-Text Transcription Services for the Bahraini Dialect

Abstract

ABSTRACT : This study explores innovative techniques to optimize the accuracy of Automated Speech Recognition (ASR) systems for the Bahraini dialect. Through the creation of custom au- dio datasets tailored for Bahraini speech, we investigated various advanced techniques to fine-tune the OpenAI Whisper model. Whisper, a multilingual speech recognition model, showed significant improvement in processing the Bahraini dialect after fine-tuning. We successfully reduced the Word Error Rate (WER) from 181.608% to 13.54% by utilizing these datasets that were specifically developed for this research.