Kaustubh Patil,Darshan Nerkar, Vishal Sharma,Sanskriti Narlawar

Speech recognition technology has seen significant advancements in recent years, enabling more efficient and accessible communication across diverse linguistic communities. In this paper, we explore the application of pretrained models for Marathi speech recognition, focusing on the "tanmaylaud/wav2vec2-large-xlsr-hindi-marathi" model provided by Hugging Face. Our study aims to evaluate the effectiveness of this pretrained model in accurately transcribing Marathi speech to text. We describe the methodology used to implement Marathi speech recognition, including data collection, preprocessing, and model selection. The experimental setup details the hardware and software environment, along with training procedures and evaluation metrics. Through experimentation and analysis, we assess the performance of the pretrained model and compare it with baseline approaches. Our findings demonstrate the viability of pretrained models for Marathi speech recognition, showcasing their potential applications in diverse fields such as accessibility tools, language learning platforms, and transcription services. We discuss the implications of our research in improving communication and technology accessibility for Marathi speakers and outline future directions for advancing Marathi speech recognition technology.

https://doi.org/10.58357/jedcs.2024.v3.i2.20