Scaling on-device spoken language understanding to new languages with large language models

Abstract Spoken language understanding (SLU) models are a core component of voice assistants (e.g., Alexa, Bixby, Google Assistant), but collecting extensive labeled data for target languages is challenging. In this paper, we introduce a data-centric pipeline to expand On-Device SLU to new languages by leveraging a large language model (LLM) for machine translation of slot-annotated English training data. The LLM is fine-tuned to preserve slot annotations during translation using an HTML tag-based slot marking strategy. Our approach is evaluated on the MultiATIS++ benchmark, a multilingual SLU dataset covering eight languages. In an On-device setting, we achieve a new state-of-the-art overall accuracy of 62.18 %, up from 55.11 % achieved by the best prior method, HC 2 L. In an Edge scenario with a tiny SLU model (5MB, no pre-training), our translated data boosts overall accuracy from a baseline 5.31 % to 22.06 %. In contrast to mentioned baselines, our LLM-based translation requires no changes to the SLU model architecture and is slot-type independent, requiring no manual slot descriptions. This work demonstrates that state-of-the-art LLMs can serve as effective “slot translators” providing a scalable path to multilingual SLU without costly SLU data collection or architecture overhaul.