Year: 2025
Type: conference-paper
Abstract: In this paper, we design a Speech-to-Text Translation (ST) system to translate English into Hindi, Bengali, and Tamil, and vice versa.We explore both cascaded and End-to-End (E2E) approaches as part of the IWSLT 2025 Indic shared task.In the cascaded systems, we leverage the pre-trained Wav2Vec2 model from AI4Bharat's Vakyansh project, and then fine-tune it for Automatic Speech Recognition (ASR).The resultant ASR outputs are then translated using the adapted IndicTrans2 Neural Machine Translation (NMT) model with IWSLT task-specific data.In the E2E approach, we train models from scratch using only the IWSLT dataset, leveraging the Fairseq Speech Translation framework which uses transformerbased encoder-decoder architecture optimized for multilingual speech inputs.In the paper, the performance of these two distinct approaches in handling low-resource Indic speech translation tasks is compared.Although in the E2E approach, the pre-trained Acoustic model is not leveraged, its results in the En-Indic setting are impressive.However, this approach does not perform well in the Indic-En setting due to lack of sufficient training data.On the other hand, the cascaded approach leverages pre-trained models and outperforms for all language pairs.
Authors: Mukund K. Roy, Karunesh Arora, Praveen Kumar Chandaliya, Rohit Kumar, Pruthwik Mishra
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4412944413