Building Collaborative Speech Corpora for Low-Resource Languages: The Galician Dataset in Mozilla Common Voice

Abstract

This paper presents the methodology and outcomes of building collaborative speech corpora in Mozilla Common Voice (MCV), focusing on the Galician case within Proxecto Nós. We describe the organization of voice collection campaigns –on-site events, student participation, Validatón marathons, and corporate collaboration– and analyze the results in MCV v22.0. While the dataset has achieved a modest scale, major gaps remain in metadata completeness and dialectal tagging, with implications for ASR performance. Drawing on our experience, we highlight effective strategies for engagement, such as transparent communication, cultural identification, and user-friendly tools. We conclude with lessons learnt for improving data representativeness, participant retention, and ethical governance. The observations are specific to the Galician case study but may inform similar efforts in other lesser-resourced languages.

Resources

Details