A TRANSFORMER APPROACH TO BILINGUAL AUTOMATED SPEECH RECOGNITION USING CODE-SWITCHED SPEECH

Abstract

In a bilingual and linguistically diverse country like India, where a significant portion of the population is fluent in multiple languages, the conventional bilingual Transformer neural network architecture faces challenges in accurately translating conversations that seamlessly switch between different languages. This limitation in translation accuracy underscores the need for more sophisticated language models. The proposed solution involves leveraging the Generative Pre- Trained Transformer (GPT) model, a powerful deep learning architecture within the transformer framework. Trained in an unsupervised manner on extensive text data, the GPT model demonstrates enhanced language understanding and generation capabilities. By pre-training on a diverse dataset, GPT learns to capture the intricacies of syntax, semantics, and contextual nuances, enabling it to accurately predict and generate coherent text. We experimented on Tamil-English data and found that the Generative Pre-Trained Transformer model can achieve an 84.37% relative accuracy rate even for short sentences and 73.98% relative accuracy rate for lengthy sentences in bilingual ASR performance. The adaptability of GPT to various downstream tasks, context-aware approach to language processing in linguistically diverse environments.