SoBERT: A Soundex Based Enhancement of Robust Speech Transformer

Automatic Speech Recognition (ASR) systems have gained significant attention due to their applications in various domains. Speech varies across regions, posing challenges for ASR and translation systems in real time. Even slight variation in pronunciation, intonation, and dialect can impact accuracy, making multilingual recognition and translation a persistent challenge particularly in case of India. This paper introduces SoBERT, a novel Soundex-based enhancement of robust speech transformer novel system, designed to reduce the Word Error Rate (WER) in speech transformer models. Additionally, it presents an analysis of an ASR system leveraging deep learning and the Transformer architecture, focusing on the OpenAI Whisper model. The SoBERT system pipeline includes the Whisper model, a probabilistic language model to identify the next most probable word and Soundex for recognizing similar words, thereby enhancing the overall accuracy of the Whisper model. Testing SoBERT on the most spoken banking everyday sentences resulted in an average reduction of 14-40% in WER compared to the existing Whisper model. The proposed SoBERT system demonstrates notable efficiency and accuracy, demonstrating the potential of the Whisper model in multilingual contexts.