Lightly supervised vs. semi-supervised training of acoustic model on Luxembourgish for low-resource automatic speech recognition

In this work, we focus on exploiting inexpensive data in orderto to improve the DNN acoustic model for ASR. We exploretwo strategies: The first one uses untranscribed data fromthe target domain. The second one is related to the proper selectionof excerpts from imperfectly transcribed out-of-domainpublic data, as parliamentary speeches. We found out that bothapproaches lead to similar results, making them equally beneficialfor practical use. The Luxembourgish ASR seed systemhad a 38.8% WER and it improved by roughly 4% absolute,leading to 34.6% for untranscribed and 34.9% for lightlysuperviseddata. Adding both databases simultaneously ledto 34.4% WER, which is only a small improvement. As asecondary research topic, we experiment with semi-supervisedstate-level minimum Bayes risk (sMBR) training. Nonetheless,for sMBR we saw no improvement from adding the automaticallytranscribed target data, despite that similar techniquesyield good results in the case of cross-entropy (CE) training.