Developing a domain-dependent Automatic Speech Recognition system with minimal data

Abstract

ABSTRACT The development of Automatic Speech Recognition (ASR) systems has substantially improved through deep learning models and extensive datasets during recent years. Accurate deployment of Automatic Speech Recognition systems within the banking domain proves difficult for the low-resource language Sinhala because of present obstacles in building efficient systems. The primary research objective of this research was to develop a domain-dependent ASR system that functions effectively with minimal data. The research investigates limitations of present-day ASR systems for low-resource languages, highlighting how few Sinhala speech datasets are accessible to the public and showing no domain-specific corpora exist. A domain-specific private dataset was developed to meet this requirement by focusing on banking-related terminology and situations. The research achieved the objectives through OpenAI's Whisper model fine-tuning process with limited data for banking applications. The Whisper model achieved performance evaluation through measurements of Word Error Rate (WER) along with loss values after receiving fine-tuning through the specified dataset. The fine-tuning process of the model produced 22.39% WER which surpassed the primary error rates of commercially available Whisper models that directly interacted with the Sinhala banking dataset. Transfer learning and fine-tuning methods play a critical role according to the research for improving ASR accuracy by using small data sets. The performance of automatic speech recognition depends on four key elements which involve data quality as well as domain suitability and model structures and appropriate tuning parameters. The research succeeded in its goals yet recognizes two major obstacles involving excessive computation requirements and insufficient database diversity. This research demonstrates two essential outcomes: first it proves the development of domainspecific Sinhala ASR systems using limited data and second it offers a framework which can apply to low-resource languages and special applications for inclusive speech recognition technology accessibility. Future research directions include expanding the dataset, optimizing computational efficiency, and exploring advanced machine learning techniques to further enhance ASR accuracy and applicability.