HW-TSC’s Submission to the IWSLT 2026 Subtitling Track

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

Summary

HW-TSC submitted a cascaded strategy for automatic subtitle generation to the IWSLT 2026 Subtitling track, designed for unconstrained conditions. Their approach begins with a large-model-based streaming speech recognition framework, which integrates VAD, sliding-window context caching, and long audio chunking. This framework utilizes the Qwen3 forced alignment model to achieve precise English speech-to-text transcription and timestamping. Subsequently, a Qwen3-based translation model performs text translation. The final stage involves ensuring compliance with subtitle constraints like characters per second (CPS) and characters per line (CPL). Segments exceeding these thresholds are quantitatively evaluated and then rewritten by a large language model, preserving semantic meaning to produce standard-compliant subtitle files.

Key takeaway

For NLP Engineers developing automated subtitling solutions, you should consider a cascaded architecture to manage complex requirements. Implementing distinct stages for speech recognition, translation, and constraint-based rewriting, as demonstrated with Qwen3 models, can significantly improve output quality and compliance. Evaluate segments quantitatively against CPS and CPL thresholds, then use a large language model to semantically preserve and adapt content. This structured approach ensures subtitle files meet industry standards efficiently.

Key insights

A cascaded approach combining speech recognition, translation, and LLM-based rewriting effectively generates compliant subtitles.

Principles

• Cascaded systems enhance complex task accuracy.
• LLMs can adapt content to strict constraints.
• Quantitative evaluation guides content adaptation.

Method

A cascaded strategy first uses a Qwen3-based streaming ASR for transcription and timestamping, then a Qwen3 translation model, and finally an LLM to rewrite segments for CPS/CPL compliance.

In practice

• Integrate VAD for robust speech processing.
• Use Qwen3 for both ASR and translation.
• Apply LLMs for constraint-driven text adaptation.

Topics

• Automatic Subtitling
• Speech Recognition
• Machine Translation
• Large Language Models
• Qwen3 Model
• IWSLT 2026

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

Related on AIssential

• How Vimeo Implemented AI-Powered Subtitles — LLMs optimized for fluency can break structural constraints, requiring architectural separation of concerns.
• Comparing Qwen3-VL AI Models for OCR Task — Quantized larger Qwen3-VL models can outperform smaller, unquantized versions in both speed and accuracy for complex OCR.
• 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.
• CAPS: Cascaded Adaptive Pairwise Selection for Efficient Parallel Reasoning — CAPS optimizes LLM parallel reasoning by adaptively reducing verifier compute through cascaded pairwise selection.
• Leveraging Audio-LLMs to Filter Speech-to-Speech Training Data — An audio-LLM can effectively filter noisy speech-to-speech translation data using a Rank-to-Distill strategy, improving S2ST performance.
• AutoAdapt: Automated domain adaptation for large language models — AutoAdapt automates LLM domain adaptation, transforming a manual, costly process into a reproducible, constraint-aware engineering discipline.

Open in AIssential →

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