MLLP-VRAIN UPV System for the IWSLT 2026 Simultaneous Speech Translation Task

2026-06-29 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, medium

Summary

The MLLP-VRAIN research group participated in the IWSLT 2026 Simultaneous Speech Translation (SimulST) track, presenting a robust cascaded solution for long-form SimulST. Their system integrates the recently released Parakeet and Qwen 3.5 models, employing adaptive black-box policies to achieve better quality-latency trade-offs. Expanding participation to all language directions, the group also engaged in the new context track for English to German, Italian, and Chinese (En→De, It, Zh). This context processing combines ASR word-boosting with a Retrieval Augmented Generation (RAG) mechanism using offline pre-translated exemplars to guide generation and enrich domain-specific context. The system demonstrated a substantial quality improvement of +5.82 XCOMET-XL on the MCIF En→De test set compared to the previous year, with context track processing further boosting performance by +1.03. A detailed latency analysis was also provided.

Key takeaway

For NLP Engineers evaluating simultaneous speech translation systems, consider integrating cascaded models like Parakeet and Qwen 3.5 with adaptive black-box policies. This approach significantly improves quality-latency trade-offs. Furthermore, explore ASR word-boosting combined with RAG mechanisms using pre-translated exemplars to inject domain-specific context, as this yielded an additional +1.03 XCOMET-XL improvement. Your next steps could involve benchmarking similar hybrid architectures.

Key insights

MLLP-VRAIN's SimulST system combines Parakeet and Qwen 3.5 with adaptive policies and RAG for improved quality and context.

Principles

• Adaptive black-box policies optimize quality-latency.
• Contextual RAG enhances domain-specific translation.
• Cascaded models improve long-form SimulST robustness.

Method

The system employs a cascaded architecture with Parakeet and Qwen 3.5, using adaptive black-box policies for SimulST. Context integration involves ASR word-boosting and RAG with pre-translated exemplars.

In practice

• Apply adaptive policies for SimulST trade-offs.
• Integrate ASR word-boosting for context.
• Use RAG with exemplars for domain specificity.

Topics

• Simultaneous Speech Translation
• IWSLT 2026
• Parakeet model
• Qwen 3.5 model
• Retrieval-Augmented Generation
• Quality-Latency Trade-offs
• ASR Word-Boosting

Best for: Research Scientist, AI Engineer, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• MLLP-VRAIN UPV system for the IWSLT 2026 Simultaneous Speech Translation task — The system combines Parakeet and Qwen 3.5 with adaptive policies and RAG for improved SimulST quality and latency.
• A Practical Evaluation Method for Long-Form Simultaneous Speech-to-Speech Translation — A new evaluation method for long-form SimulS2ST uses ASR, forced alignment, and sentence embeddings to measure latency and quality.
• HARP: Hadamard-Preconditioned Adaptive Rotation Processor for Extreme LLM Quantization — HARP adaptively rotates LLM weights for extreme low-bit quantization, improving accuracy and inference speed over fixed methods.
• Multilingual Reasoning Cascades Need More Context — Multilingual reasoning cascades improve by preserving original query context for final translation.
• Hierarchical Policy Optimization for Simultaneous Translation of Unbounded Speech — HPO improves simultaneous speech translation by post-training models with a hierarchical reward that prioritizes quality over latency.
• Alibaba Qwen Team Introduces Qwen3.5-LiveTranslate-Flash: Real-Time Multimodal Interpretation Across 60 Languages at 2.8-Second Latency — Alibaba's Qwen3.5-LiveTranslate-Flash delivers real-time, multimodal interpretation across 60 languages with 2.8-second latency and voice cloning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.