Amulwe Kimün: A Community-Grounded Demo, Resource, and ASR Baseline for Mapuzugun

Abstract

This paper introduces Amulwe Kimün ("a means or path for knowledge" in Mapuzugun), a community-grounded multimodal quiz application co-created with Mapuche speakers to support the revitalization of Mapuzugun. Developed within a FACSO–CONADI collaboration during an intensive language course, the platform integrates multiple-choice, ordering and free-text exercises, as well as forums and chat functions to promote language practice, peer learning, and a sense of community. A pilot involving 32 learners produced 562 responses across 43 questions, with accuracies of 92.3% (multiple choice), 55.2% (ordering), and 7.1% (free-text), offering insights for refining item design and evaluation strategies. The low open-answer accuracy is related to a strict exact-match scoring and orthographic variation of the language. In addition, we present an initial Automatic Speech Recognition (ASR) prototype (Whisper-small + LoRA), establishing a fine-tuned baseline relative to zero-shot performance. The demo illustrates how community-grounded design, language resources, and lightweight evaluation can productively meet in a practical tool for an endangered language.

Resources

Details