Abstract
Large Language Models (LLM) have revolutionised natural language processing and its applications. However, high-performance LLMs require copious data and computing resources for their development and
are rarely public. This also concerns Large Acoustic Models (LAM) for processing spoken language. The Phoné initiative seeks to build an open Italian speech dataset to advance Automatic Speech Recognition (ASR) systems and support public research. Spearheaded by institutions in Naples, Pisa, and Bolzano, the project gathers diverse Italian audio sources and applies advanced ASR architectures, including supervised and self-supervised models. This paper details Phoné’s dataset creation, ASR model evaluation, and ethical considerations, aiming to democratise access to Italian-language resources and foster innovation in ASR technologies.