End-to-End Speech Recognition for Endangered Languages of Nepal

AbstractThis paper presents three experiments to test the most effective and efficient ASR pipeline to facilitate the documentation and preservation of endangered languages, which are often extremely low-resourced. With data from two languages in Nepal —Dzardzongke and Newar— we show that model improvements are different for different masses of data, and that transfer learning as well as a range of modifications (e.g. normalising amplitude and pitch) can be effective, but that a consistently-standardised orthography as NLP input and post-training dictionary corrections improve results even more.