The Eloquence team has detailed their submission for the Interspeech 2026 MLC-SLM challenge, focusing on multilingual multiple-choice question answering across 21 languages. They explored three distinct methods: fine-tuning Voxtral-Mini-3B with LoRA and various data augmentations, employing multimodal in-context learning with the Voxtral-24B model to mitigate label bias, and developing a training-free retrieval system utilizing a voice-anchored memory. The in-context learning approach yielded the best performance with a macro-accuracy of 0.72, while all three systems surpassed the challenge's baseline. AI
IMPACT This research contributes to advancements in multilingual question-answering systems and explores novel techniques for improving model performance on diverse language tasks.
RANK_REASON The item is a research paper detailing a submission to a challenge. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
