Eloquence improves multilingual audio quiz answering accuracy

The Eloquence submission for Task 2 of the Interspeech 2026 MLC-SLM challenge

Computation and Language

Summary

Answering multiple-choice questions from audio in many languages is hard because of differences in speech and text. The Eloquence team tried three ways to solve this: changing a smaller model with added language variation and audio tweaks, using a big fixed model to fix mistakes by looking at examples, and a system that remembers sounds, meanings, and facts without training. Their best method reached 81% accuracy, beating the existing baseline by a lot.

What this means in practice

Authors

Jordi Luque, Lorenzo Concina, Marco Matassoni, Alessio Brutti, Filippo Vella

Abstract

This paper details the Eloquence team's approach to Task 2 of the 2nd MLC-SLM challenge at Interspeech 2026, which involves multilingual Multiple-Choice Question Answering (MCQA) across 21 languages. Three approaches are explored. First, we fine-tune Voxtral-Mini-3B via LoRA with cross-lingual data augmentation, ASR transcript augmentation and timestamp-aware audio cropping, achieving 0.72 macro-accuracy on evaluation Phase 2. Second, we apply multimodal in-context learning (ICL) to the frozen Voxtral-24B model to correct a strong label bias, reaching 0.81, our best result. Third, a training-free retrieval system based on a three-layer voice-anchored memory combining acoustic identity, semantic content, and a knowledge graph achieves 0.68. All three systems substantially outperform the official baseline.