Shahzia Sayyad

For Automatic Speech Recognition (ASR) and Natural Language Processing (NLP) systems, code-switching—the habit of alternately speaking in several languages within a single conversation—offers special difficulties and possibilities. ASR systems have to efficiently manage language transitions as multilingual communication gets more common if we want real-time speech recognition. This work investigates innovative approaches for processing code-switched audio, solves the dearth of multilingual datasets, and assesses several technologies applied to identify and analyze mixed-language speech. Emphasizing Hindi-Marathi code-switching, we present a dynamic language-switching architecture leveraging reinforcement learning methods including Q-Learning and Deep Q-Networks (DQN) to improve language transition identification. Moreover, we present a dataset especially meant for multilingual voice recognition and evaluate ASR performance with Character Error Rate (CER) and Word Error Rate (WER). Our study reveals current constraints and provides future directions to improve ASR adaptation, therefore guaranteeing more accurate and strong recognition in many multilingual settings.