Publication

Abstract : India has a diverse verbal landscape, comprising 22 major languages and approximately 1600 dialects. It is essential to recognise and comprehend the distinct features of each language and accent as it helps in promoting effective communication, learning sociolinguistic phenomena and language learning. This study, focuses especially on distinguishing between Hindi and non-Hindi languages. Unlike previous works, the dataset used in this work is newly curated, offering a fresh perspective on dynamic audio processing. Accent detection is made accurate by utilising audio preprocessing, which extracts Mel-frequency cepstral coefficients (MFCCs) and then runs them through convolutional neural networks (CNN). This process adds variation to the understanding of the diversity of languages. A standard existing MobileNet model demonstrates superior performance as compared to other existing models with an accuracy of 82%. Additionally, LIME (Local Interpretable Model-agnostic Explanations) analysis is applied to understand the model’s predictions, providing insights into the regions of the spectrograms influencing the classification. Through this innovative method, accent barriers are not only removed but also transformed into bridges, facilitating improved communication, enhanced educational opportunities, and promoting a more connected India.