Multilingual Long-Form Speech Instruction Following: KIT's Submission to IWSLT 2026

KIT has submitted a multilingual speech instruction following approach to IWSLT 2026, combining data augmentation and LLM-based label generation to yield over 1M instances across six tasks and four languages^[1].

The IWSLT Instruction Following Track introduced new tasks this year, including an unknown surprise task, posing a challenge against overfitting to known tasks^[1]. KIT's submission addresses this challenge by using a general data augmentation pipeline that converts short-form corpora into long-form training data. The approach also incorporates cross-lingual translation and LLM-based label generation. In related work, researchers have focused on cross-lingual voice cloning, a task central to speech translation^[2]. A key challenge in this area is maintaining intelligibility and naturalness in the presence of accent variation and domain-specific vocabulary^[2]. To address this, techniques such as language tag prompting and reinforcement learning fine-tuning have been introduced to improve language control and intelligibility^[2]. Additionally, a reference-conditioned lexical matching method has been proposed to improve pronunciation of domain-specific terms^[2].