Improving Automated Speech Recognition Using Retrieval-Based Voice Conversion

Anas Mohammed Alhumud, Muhammad AL-Qurishi, Yasser Omar Alomar, Ali Alzahrani, Riad Souissi

Keywords: Voice Conversion Techniques; Automatic Speech Recognition (ASR); Non-Native English Speakers; OpenAI Whisper Models; Word Error Rates (WER)

Abstract: This study examines the efficacy of voice conversion techniques in enhancing Automatic Speech Recognition (ASR) accuracy for non-native English speakers. Utilizing the OpenAI Whisper models, we analyzed transcription accuracy across various accents and countries. Significant reductions in Word Error Rates (WER) were observed, with the Whisper Large-v2 model showing the most pronounced improvements. Our findings indicate that advanced voice conversion can mitigate accent bias, promoting inclusivity and broadening the applicability of ASR technology to a more diverse user base.

Submission Number: 103