Recent Trends in Speech Recognition for Diverse Indian Languages

Language inclusivity has become a prevailing trend in the digital landscape, yet the realm of speech technology for Indian languages remains relatively underexplored compared to English. This paper aims to bridge this gap by building a common platform for advancing ASR (Automatic Speech Recognition) technology for diverse Indian languages. Various training architectures have been investigated, employing a diverse array of methodologies during the training phase, including WeNet, Espnet, K2, and NeMo. These architectures were applied across speech corpus ranging from low-resource to large datasets, encompassing both open-source and proprietary corpora. Ultimately, the architecture that demonstrated superior performance in terms of accuracy and Word Error Rate (WER) was chosen for each language. Through this study, the authors seek to contribute to the broader objective of building a robust ASR system promoting language inclusivity in the digital sphere, thereby facilitating more accessible and inclusive communication experiences for diverse linguistic communities.