AlignAtt4LLM: Fast AlignAtt for Decoder-Only LLMs at 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

AlignAtt4LLM is an IWSLT 2026 simultaneous speech translation system designed for English to German, Italian, and Chinese. This system employs a synchronous cascade, utilizing Qwen3-ASR for forced alignment to produce incrementally updated source transcripts, which are then translated by Gemma-4 E4B-it under an MT-side AlignAtt policy. Notably, this marks the first application of AlignAtt to a decoder-only Large Language Model, addressing the absence of traditional encoder-decoder cross-attention. The system achieves a usable policy through four key innovations: an explicit source span in the prompt, offline selection of translation-specific alignment heads, selective qk-fast replay of the draft-to-source attention block, and runtime query/key capture that preserves model outputs bit-identically. On the IWSLT 2026 development set, AlignAtt4LLM surpassed baselines for English to German and Italian in both 2-second low-latency and sub-4-second CU-LongYAAL high-latency scenarios, though English to Chinese results were mixed. The approach is adaptable to other decoder-only MT backbones.

Key takeaway

For NLP Engineers developing simultaneous speech translation systems, this work demonstrates a viable path for integrating AlignAtt with decoder-only LLMs. You should consider adapting the proposed four-step policy, including explicit prompt source spans and selective attention head replay, to utilize models like Gemma-4 E4B-it. This approach offers improved latency and accuracy for European languages, and its generalizability means you can apply it to other strong MT backbones for diverse target languages.

Key insights

AlignAtt4LLM adapts AlignAtt for decoder-only LLMs, achieving competitive simultaneous speech translation performance.

Principles

• AlignAtt can be adapted for decoder-only LLMs.
• Explicit prompt spans guide attention.
• Selective attention head calibration improves MT.

Method

A synchronous cascade uses Qwen3-ASR for ASR and Gemma-4 E4B-it for MT. It employs explicit prompt source spans, offline alignment head selection, selective qk-fast replay, and runtime query/key capture for AlignAtt.

In practice

• Apply AlignAtt to decoder-only LLMs.
• Use Qwen3-ASR with Gemma-4 E4B-it.
• Reapply policy to other MT backbones.

Topics

• Simultaneous Speech Translation
• Decoder-Only LLMs
• AlignAtt
• Qwen3-ASR
• Gemma-4 E4B-it
• Attention Mechanisms

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

Related on AIssential

• 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.
• Matching Tasks to Objectives: Fine-Tuning and Prompt-Tuning Strategies for Encoder-Decoder Pre-trained Language Models — Aligning task requirements with pre-training objectives significantly boosts encoder-decoder PLM performance.
• Align Once, Benefit Multilingually: Enforcing Multilingual Consistency for LLM Safety Alignment — A plug-and-play MLC loss improves multilingual LLM safety alignment using only prompt variants, reducing resource needs.
• Gemma 4 12B MTP Local Test | Coding, OCR, Visual RAG with llama.cpp — The Gemma 4 12B MTP model offers multimodal capabilities and efficiency, nearing larger MoE models, but with performance trade-offs.
• Montreal Forced Aligner and the state of speech-to-text alignment in 2026 — MFA 3.0 achieves leading speech-to-text alignment performance through continuous development and advanced linguistic features.
• Streaming Speech-to-Text Translation with a SpeechLLM — A novel SpeechLLM adaptively decides when to emit translation tokens, achieving low latency and high quality.

Open in AIssential →

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