Cascaded Modular or End-to-End? : An Investigation on Speech-to-Speech Translation Task for Dravidian Languages
Year: 2026
Type: conference-paper
Abstract: This paper presents a study of speech-to-speech translation for low-resource Dravidian languages, focusing on Tamil, Telugu, and Kannada.We investigate the efficacy of the Cascaded Modular system with the End-to-End system in both zero-shot and fine-tuned settings.The Cascaded Modular approach combines an ASR Module (Whisper Large-v3 for English speech; IndicConformer for Dravidian speech), a Text-to-Text translation module (In-dicTrans2), and a Speech synthesis module (Indic Parler-TTS), whereas SeamlessM4T-v2 is used as the End-to-End system.For parameterefficient Low-Rank Adaptation (LoRA) finetuning to adapt the translation component to domain-specific datasets, we use FLEURS and Mann-ki-Baat (a subset of the BhasaAnuvaad dataset).Cascaded Modular systems achieve BLEU scores ranging from 3.17 to 18.96 in the zero-shot setting and 5.08 to 19.18 after fine-tuning, whereas the End-to-End model ranges from 3.02 to 15.72 in zero-shot settings across languages and 4.11 to 16.84 after finetuning.The results show that Cascaded Modular systems generally outperform the End-to-End model across most setups, though the margin varies across language pairs.Parameterefficient fine-tuning yields notable improvements in translation quality and speech generation performance for low-resource Dravidian speech translation.The code repository is made publicly available 1 .
Authors: Bhavana Nali, Abhik Jana
Institution: Indian Institute of Technology Bhubaneswar
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166824292