Data Augmentation and Subword based Approach for Improved Speech Processing in Low Resource Languages

Abstract

Communication among humans is predominantly through speech. The ubiquity of esmartphones necessitates making speech technology applications such as voice assis- newlinetants reachable to users in all different regions of the world. Given the fact that speech newline newlineis a natural interface and a large percentage of the Indian population is semi-literate, newline newlinespeech is the preferred interface for mobile phone users in India. Therefore, extend- newlineing the benefits of speech-processing applications, such as voice assistants in Indian newline newlinelanguages to smartphone users is important. This warrants the need to build speech newlinecorpora for every language spoken by the multilingual community of India, which in newlineturn involves the tedious task of obtaining speech recordings from different speakers newline newlineand annotating the recordings with corresponding text. Hence, a majority of the In- newlinedic languages are low-resourced, having only tens of hours of annotated speech. This newline newlinemakes research in speech processing for low-resource languages very relevant. newline