Empowering Low-Resource Language Translation: Methodologies for Bhojpuri-Hindi and Marathi-Hindi ASR and MT
Harpreet Singh Anand, Amulya Ratna Dash, Yashvardhan Sharma · 2024
This paper presents the methodologies implemented for the Automatic Speech Recognition and Machine Translation for the language pairs Bhojpuri-Hindi and Marathi-Hindi for the Dialectal and Low-Resource shared task proposed by The International Conference on Spoken Language Translation (IWSLT) for 2024.The implemented method uses the transcriptions generated through a fine-tuned Whisper models(for Marathi-Hindi) and vakyansh-wav2vec model (for Bhojpuri-Hindi) and generates the translations using fine-tuned NLLB(No Language Left Behind) Models for both the tasks.The selection of more accurate translation is done through sentence-embeddings generated using the MuRIL(Multilingual Representations for Indian Languages) (Khanuja et al., 2021) model for the Marathi-Hindi task.