Bilibili Open-Sources Index-Translate, a Multilingual Translation Model Family Based on Qwen3.5
Bilibili has open-sourced Index-Translate, a multilingual model family built on Qwen3.5 that supports text translation across 150 languages with strict formatting and terminology compliance. The suite extends beyond basic text to offer specialized models for voice-preserved speech translation, syllable-constrained output, and full-document context translation. By open-sourcing models tailored for real-world multimedia localization—such as voice preservation and syllable matching for dubbing—Index-Translate lowers the barrier for automated video subtitling and dubbing. It provides media creators and developers with open-weights tools that address practical translation constraints often missed by general-purpose LLMs. The release features four specialized components: Index-Translate for text and structured content, Index-Echo for subtitles and speech conditioned on the source speaker's voice, Index-Homura for adjusting target syllable counts, and Index-NativeLong for cross-passage document translation.
## BACKGROUND
Machine translation has traditionally focused on literal accuracy, but tasks like video dubbing require syllable-constrained translation to ensure target text fits the duration and lip movements of original dialogue. Additionally, voice-preserved speech translation enables direct speech-to-speech transfer while maintaining speaker identity, pitch, and timbre without full voice-cloning pipelines.