HW-TSC’s Submissions to the IWSLT 2026 Offline Speech Translation Task

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

Summary

HW-TSC submitted an advanced speech translation system to the IWSLT 2026 Offline Speech Translation Task, targeting English-to-Chinese and English-to-German unconstrained tracks. The system employs a robust cascade architecture specifically optimized for processing long-form, unsegmented audio. To address common issues like hallucination and inconsistency in long-sequence processing, it implements a two-pass transcription strategy. This involves an initial streaming ASR with a 12-second context buffer for sentence coherence, followed by Qwen3-ForcedAligner for precise timestamping. A second refinement pass then uses Qwen3-Omni on re-segmented 30-second chunks to achieve high-fidelity transcriptions. For translation, a context-aware segment merging strategy, handling up to 150 tokens, provides the Qwen3 LLM with sufficient semantic context. Benchmarking on tst-2022 showed COMET scores of 0.8462 for English-to-Chinese and 0.7854 for English-to-German, significantly surpassing standard cascade baselines.

Key takeaway

For NLP Engineers developing robust speech translation systems, consider adopting a multi-stage approach for long audio. Your systems can mitigate hallucination and inconsistency by implementing a two-pass transcription strategy, leveraging precise timestamping and chunk-based refinement. Additionally, providing large language models with context-aware merged segments, up to 150 tokens, will significantly enhance translation fidelity, as demonstrated by the 0.8462 (En-Zh) and 0.7854 (En-De) COMET scores.

Key insights

A two-pass transcription and context-aware merging strategy improves long-form speech translation quality.

Principles

• Cascade architecture is effective for speech translation.
• Long-sequence processing needs hallucination mitigation.
• Semantic context improves LLM translation performance.

Method

A two-pass transcription uses streaming ASR with a 12-second buffer, Qwen3-ForcedAligner for timestamping, and Qwen3-Omni on 30-second chunks for refinement. Translation merges segments up to 150 tokens.

In practice

• Implement two-pass ASR for long audio.
• Use Qwen3 models for transcription and translation.
• Merge segments up to 150 tokens for LLM context.

Topics

• Speech Translation
• Cascade Architecture
• Automatic Speech Recognition
• Large Language Models
• Qwen3
• Long-Form Audio Processing
• IWSLT 2026

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• FBK's Long-form SpeechLLMs for IWSLT 2026 Instruction Following — Fixed 30-second segmentation provides robust long-form SpeechLLM performance, mitigating hallucination effects on ASR and SSUM.
• Multilingual Reasoning Cascades Need More Context — Multilingual reasoning cascades improve by preserving original query context for final translation.
• Streaming Speech-to-Text Translation with a SpeechLLM — A novel SpeechLLM adaptively decides when to emit translation tokens, achieving low latency and high quality.
• Decoding LLMs — Part 2: A Step-by-Step Journey Into the Mind of Modern AIe — The Transformer decoder translates encoder-generated contextual representations into a target language using a multi-layered stack.
• Analyzing Error Propagation in Korean Spoken QA with ASR-LLM Cascades — ASR-stage information loss, especially single-character errors in Korean, is the primary bottleneck for ASR-LLM cascade performance.
• Understanding LLM Behavior in Multi-Target Cross-Lingual Summarization — LLMs perform cross-lingual summarization and translation jointly in later layers, not sequentially.

Open in AIssential →

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