A First LVCSR System for Luxembourgish, a Low-Resourced European Language
Résumé
Luxembourgish is embedded in a multilingual context on the divide between Romance and Germanic cultures and remains one of Europe’s low-resourced languages. We describe our efforts in building a large vocabulary ASR system for such a “minority” language without resorting to any prior transcribed audio training data. Instead, acoustic models are derived from major European languages. Furthermore, most Luxembourgish written sources include significant parts in other languages. This poses specific challenges to Language Model estimation. Some scientific and technological issues addressed include: (i) how to build acoustic models if no labeled acoustic training data are available for the under-resourced target language? (ii) how to make use of the new system to accelerate resource production for the target language? (iii) how to build a vocabulary and a language model with multilingual written texts? (iv) how to determine the “best” phonemic inventory for ASR? First ASR results illustrate the accuracy of the various sets of monolingual and multilingual acoustic models and what these suggest concerning language typology issues.