Modélisation acoustico-phonétique de langues peu dotées : Études phonétiques et travaux de reconnaissance automatique en luxembourgois

Luxembourgish, a Germanic-Franconian language, is embedded in a multilingual context on thedivide between Romance and Germanic cultures and remains one of Europe’s under-describedlanguages. This paper investigates the similarity between Luxembourgish phone segments withGerman, French and English via forced speech alignment techniques. Making use of monolingualacoustic seed models from these three languages, as well as “multilingual” models trained onpooled speech data we investigated whether Luxembourgish was globally better represented byone of the individual languages or by the multilingual model. While globally, the German modelsprovide the best match, a phone-based analysis, shows language-specific preferences. First ASRresults illustrate the accuracy of the various sets of monolingual and multilingual acoustic modelsand Luxemburgish acoustic models built from 1200 hours of untranscribed Luxemburgish audiodata using unsupervised methods