SardinianVoxes: A Speech Recognition Dataset for the Sardinian Languages

Low-resource languages inherently lack the extensive data typically required to train effective intelligent systems, leading to a substantial gap in technological support. In this paper, we address this challenge by focusing on the language varieties spoken in Sardinia, which are marked by limited and fragmented linguistic resources. First, we present the design and implementation of a reproducible pipeline that led to the preparation of SardinianVoxes, an audio-text dataset comprising approximately 170 hours of transcribed speech, carefully annotated to reflect the internal linguistic diversity of Sardinian. Second, we introduce a tailored evaluation protocol and benchmarking suite aimed at assessing the performance of state-of-the-art speech-to-text models, and their fine-tuned counterparts, across Sardinian varieties. We expect that our contributions will support the development of intelligent speech technologies for Sardinian.