SautiLearn: Improving Online Learning Experience with Accent Translation

Our work addresses the problem of accent translation, for the modification of spoken audio into Nigerian accents. In this work, we propose a unique speech dataset, SautiDB, consisting of 919 voices from a mixture of different Nigerianaccents (Yoruba, Igbo, Hausa, Efik-Ibibio, Igala, Edo) collected from a distributed network of volunteer speakers. We show that by using phonetic posteriorgrams in a sequence-to-sequence model, we can achieve convincing performance in accent conversion. We intend to use our resulting model to convert the accents of lectures provided in American English to a mixture of Nigerian accents that are more easily understood by native speakers of the Nigerian language. The end result is will be a proposed tool, SautiLearn.