Automatic speech recognition for low-resource languages: exploring transfer learning from related languages

Abstract

Automatic speech recognition systems typically require huge amounts of data, which are not always available for languages lacking digital resources. In this work we address the problem of automatic speech recognition for low-resourced languages by apply transfer learning approaches to the large-scale pre-trained model wav2vec2-xls-r. In particular, we present a sequential adaptation strategy where, using a common tokenizer, the large-scale model is firstly fine-tuned on a genealogically related well-resourced language (“pivot”) and then further adapted on the target low-resourced one. Experimental results in terms of word error rate, show that this approach considerably outperforms models directly fine-tuned on the target language, in particular when very few data are available. Experiments involve Italian, Spanish, Portuguese and German as well-resourced languages and two minority languages: Galician and the variety of Romansh called Vallader.