South Tyrolean Dialect-to-Standard Speech Translation: A Resource

Abstract

This paper presents a developing oral resource for South Tyrolean, a German dialect spoken in Northern Italy. The dialect is ubiquitous in spoken communication but lacks a standardised orthography. In this context, strict transcription into dialect is of limited to no utility to the local community. Instead, there is a distinct and strong demand for technology capable of directly translating spoken dialect into Standard German. To address this specific need, we introduce a dynamic, incrementally growing dataset designed to fine-tune ASR models for this translation task. Our corpus aggregates diverse sources, including media and research interviews, totalling over 13 hours of aligned audio. We describe a collaborative workflow where community partners contribute audio archives in exchange for automated transcriptions, creating a virtuous cycle of data improvement. Additionally, we detail our iterative model fine-tuning strategy, data collection challenges and the resulting improvements in model performance.

Resources

Details