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

2026-06-15 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

The MLLP-VRAIN research group submitted its system for the IWSLT 2026 Simultaneous Speech Translation (SimulST) task, utilizing the recently released Parakeet and Qwen 3.5 models. This cascaded solution addresses long-form SimulST through adaptive "black-box" policies, with explorations into policy relaxations to optimize quality-latency trade-offs. The system participates in all language directions, and for En→{De, It, Zh}, it includes a new context track. This context track integrates ASR word-boosting and a Retrieval-Augmented Generation (RAG) mechanism, leveraging offline pre-translated exemplars to guide generation and enrich domain-specific context. A detailed latency analysis is also provided. 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 processing adding another +1.03 improvement.

Key takeaway

For NLP Engineers developing simultaneous speech translation systems, consider integrating adaptive "black-box" policies with models like Parakeet and Qwen 3.5. Your team should explore policy relaxations to fine-tune quality-latency trade-offs. Additionally, for domain-specific contexts, implement ASR word-boosting alongside a RAG mechanism using pre-translated exemplars to significantly enhance performance, as demonstrated by the +5.82 XCOMET-XL improvement.

Key insights

The system combines Parakeet and Qwen 3.5 with adaptive policies and RAG for improved SimulST quality and latency.

Principles

• Adaptive "black-box" policies enhance SimulST.
• RAG with exemplars improves domain context.
• Policy relaxation balances quality and latency.

Method

A cascaded SimulST solution uses Parakeet and Qwen 3.5, applying adaptive "black-box" policies. The context track employs ASR word-boosting and RAG with pre-translated exemplars.

In practice

• Implement adaptive policies for SimulST.
• Use ASR word-boosting for context.
• Integrate RAG with pre-translated data.

Topics

• Simultaneous Speech Translation
• Parakeet
• Qwen 3.5
• Adaptive Policies
• Retrieval-Augmented Generation
• Quality-Latency Trade-offs

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Engineering Qwen-VL for Production: Vision Module Architecture and Optimization Practices — Optimizing Qwen-VL's vision module on AMD MI308X GPUs significantly boosts multimodal inference performance through kernel-level enhancements.
• AlignAtt4LLM: Fast AlignAtt for Decoder-Only LLMs at IWSLT 2026 Simultaneous Speech Translation Task — Adapting AlignAtt for decoder-only LLMs enables simultaneous speech translation by re-engineering attention mechanisms.
• A Comparative Study of Language Models for Khmer Retrieval-Augmented Question Answering — RAG efficacy for low-resource languages like Khmer is bottlenecked by retriever performance, with generator strengths varying across RAGAS metrics.
• Translate-R1: Cost-Aware Translation Tool Use via Reinforcement Learning — Reinforcement learning enables LLMs to intelligently decide when to use translation tools, optimizing cost and performance.
• System Report for CCL25-Eval Task 5: New Dataset and LoRA-Fine-Tuned Qwen2.5 — Domain-specific LoRA fine-tuning on Qwen2.5-14B with CCPoetry-49K significantly enhances classical Chinese poetry understanding and translation.
• Now in Foundry: Qwen3-Coder-Next, Qwen3-ASR-1.7B, Z-Image — New Hugging Face models offer efficient, high-performance AI for coding, multilingual ASR, and advanced image generation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.