DOA: Training-Free Decoder-Only Attention Policy for Long-Form Simultaneous Translation with SpeechLLMs

Researchers introduce DOA (Decoder-Only Attention), a training-free method that enables simultaneous speech-to-text translation using decoder-only SpeechLLMs by extracting alignment signals from self-attention mechanisms. The approach achieves low-latency, long-form translation quality comparable to offline decoding without requiring model retraining.

The research addresses a fundamental challenge in real-time speech translation: how to decide when to process input versus generate output during live streaming. Traditional encoder-decoder models use cross-attention to explicitly align source and target languages, but newer SpeechLLMs rely exclusively on self-attention, making their alignment capabilities unclear.

This work builds on the broader trend toward unified, decoder-only architectures in large language models. As foundation models consolidate around self-attention-only designs for efficiency and scaling, researchers must adapt real-time processing pipelines to these new constraints. The paper validates this approach on production-grade models like Phi4-Multimodal and Qwen3-Omni, demonstrating practical applicability.

The significance lies in its training-free nature. Most streaming translation systems require expensive fine-tuning or rely on simplistic heuristics like wait-k policies that ignore linguistic context. DOA extracts alignment information from existing model weights through clever attention analysis, eliminating retraining overhead. This lowers barriers for deploying simultaneous translation across diverse language pairs and resource-constrained environments.

The ability to handle long-form translation matters for real-world scenarios like conferences, live broadcasts, and multilingual meetings where maintaining quality over extended durations is critical. The findings suggest decoder self-attention contains richer alignment information than previously assumed, opening new research directions for understanding how these models encode structural relationships implicitly.

• →DOA extracts alignment signals from self-attention in decoder-only SpeechLLMs without requiring model retraining.
• →The method achieves near-offline translation quality while maintaining low latency for simultaneous speech-to-text translation.
• →Training-free approaches reduce deployment friction and make real-time translation more accessible across different models.
• →Decoder self-attention contains sufficient structural information to guide streaming translation policies effectively.
• →Long-form translation capabilities enable practical applications in live conference interpretation and multilingual broadcasting.