Digital Preservation of Aromanian Through Knowledge Management and Automatic Speech Recognition Evaluation

Abstract

This paper presents a knowledge management framework for the digital preservation of Aromanian, an endangered Eastern Romance language spoken across the Balkans, combined with the first systematic evaluation of automatic speech recognition (ASR) models on Aromanian dialectal speech. The proposed three-module framework encompasses localization of existing resources, distribution through digital platforms, and creation of new linguistic content through technological innovation. Empirical analysis of knowledge management practices among 176 respondents validates the framework design, revealing that information quality (28.3%) and personal utility (24.1%) are the most valued criteria for knowledge sharing, while digital information seeking (27.1%) is the dominant behaviour. Within this framework, we evaluate models from the OpenAI Whisper family across three sizes (medium, large-v2, large-v3) and multiple language settings on two Aromanian varieties: Gramosteanj and Crushova. All configurations yield word error rates (WER) above 88%, with character error rates (CER) as low as 34% under optimal conditions, indicating partial phonotactic capture despite word-level failure. The Latin language setting with large-v3 consistently achieves the best results. Romanization of non-Latin script output substantially reduces CER, confirming script mismatch as a major error source. These findings underscore the limitations of current pretrained ASR models for endangered languages and the need for dedicated resources and adaptation strategies within a broader language preservation framework

Resources

Details