Designing Matched and Mismatched ASR Systems for Punjabi Language Using DNN: Enhancing Performance Through Feature Extraction and Comparative Analysis

Abstract

In this paper, we present an innovative approach to designing Automatic Speech Recognition (ASR) systems for the Punjabi language using Deep Neural Networks (DNN). The primary objectives of this research are to build accurate and robust ASR systems, improve performance through the extraction of key features, and design matched and mismatched systems that address linguistic diversity. Our approach leverages advanced feature extraction techniques to identify critical acoustic and linguistic features that significantly impact system performance. We explore the design and implementation of both matched and mismatched ASR systems, employing DNN architectures to enhance recognition accuracy in diverse scenarios. Furthermore, we conduct a comprehensive comparative analysis between the baseline ASR systems and our proposed models to demonstrate the efficacy of our methods. The results indicate a substantial improvement in recognition accuracy, highlighting the potential of our approach to overcome challenges in ASR for underrepresented languages like Punjabi. This research contributes to the field of speech recognition by providing a framework for building robust ASR systems tailored to specific linguistic contexts, paving the way for further advancements in language-specific ASR technologies.