EDUSTT: In-Domain Speech Recognition for Nigerian Accented Educational Contents in English

Abstract

English Automatic Speech Recognition systems are trained on regular speech, therefore they may struggle to perform well on accented and domain-specific speech. For broader applications of ASR systems, such as in education, where there is synchronous learning, it is important to have a reliable system - a specialized system that recognises terms used in school subjects and spoken by accented teachers. English is our official language in Nigeria, and it is the major language used to teach in schools. However, our teachers hail from different parts of the country, where their mother-tongue affects the way they pronounce certain words. The aim of this paper is to propose an ASR system for education in Nigerian accent. Our experiment leveraged on fine tuning NeMo’s QuartzNet15x5 English model on our accented educational data. This process yielded a WER of 27\%.