MLLP-VRAIN UPV system for the IWSLT 2026 Simultaneous Speech Translation task

Short summary

MLLP-VRAIN presents a cascaded simultaneous speech translation system for IWSLT 2026 using Parakeet and Qwen 3.5, achieving +5.82 XCOMET-XL quality improvement over prior year on En→De. The system employs adaptive policies and context-aware processing with ASR word-boosting and RAG-guided generation. Additional +1.03 gains reported on the context track with offline exemplar retrieval.

• •Uses Parakeet and Qwen 3.5 models in cascaded architecture with adaptive policies for quality-latency trade-offs
• •Achieves +5.82 XCOMET-XL improvement vs. prior year; +1.03 additional gain leveraging new context track
• •Employs ASR word-boosting and RAG with offline pre-translated exemplars to guide generation and inject domain context

Generated with AI, which can make mistakes.