MEUSLI: Open-source multilingual projector brings ASR to 28 European languages
First open-science projector linking Whisper to multilingual LLMs for 28 languages
MEUSLI, introduced by Lorenzo Concina, Seraphina Fong, Marco Matassoni, and Alessio Brutti, is the first open-science multilingual projector family designed to link a Whisper speech encoder with open-source multilingual large language models (LLMs). The system enables fully open-source end-to-end automatic speech recognition (ASR) across 28 European languages, including both high- and low-resource languages. By using proper continual learning techniques, MEUSLI can be easily extended to new languages not seen during initial training, making it a scalable foundation for multilingual speech understanding.
Beyond ASR, MEUSLI demonstrates versatility in tasks like multilingual speech translation and topic identification, requiring only a few hours of task-specific supervision per language. This work addresses the limitation of existing lightweight projectors that typically support only a few languages, often restricted to English. The approach is fully open-source, promoting inclusive and reproducible research in SpeechLLMs. The paper is available on arXiv (2607.22100) under a permissive license, with citations and code expected to follow.
- First open-science multilingual projector linking Whisper with multilingual LLMs for ASR
- Supports 28 European languages, including low-resource ones, with easy extension via continual learning
- Also enables speech translation and topic identification with only a few hours of supervision per language
Why It Matters
Democratizes multilingual ASR by providing a scalable, open-source framework that works for many languages.