We present low-resource Bhojpuri-Hindi speech translation systems for the IWSLT 2026 shared task, covering both end-to-end and cascaded settings. Our end-to-end model connects a Bhojpuri-finetuned Wav2Vec2 encoder to a pretrained NLLB-200 decoder via a lightweight interconnection adapter that combines learnable layer aggregation, CNN-based temporal compression, and Transformer refinement, with optional LoRA-based decoder adaptation. For our cascaded system, we finetune Whisper for Bhojpuri ASR and NLLB-200 for Hindi MT, and further apply QE Fusion with COMET-Kiwi to improve translation selection from beam candidates.

This system description paper presents the details of our primary and contrastive approaches to translating Maltese into English for IWSLT 24. The Maltese language shares a large vocabulary with Arabic and Italian languages, thus making it an ideal candidate to test the cross-lingual capabilities of recent state-of-the-art models. We experiment with two end-to-end approaches for our submissions: the Whisper and wav2vec 2.0 models. Our primary system gets a BLEU score of 35.1 on the combined data, whereas our contrastive approach gets 18.5. We also provide a manual analysis of our contrastive approach to identify some pitfalls that may have caused this difference.