Investigating Issues with Machine Learning for Accent Classification

Abstract

Speech recognition has become a widely researched topic for decades, and there were already some successful products which has been put into commercial use, like Siri. However, sometimes it is hard to distinguish the word because words of different accents have different pronunciations. For instance, in Japanese English /r/ is usually pronounced as /l/. Therefore, it is natural to think that speech recognition could be divided into two parts. First attach a label to the audio about the accent, then recognize the contents based on the regular pattern of that accent. In this paper, we researched on several characteristics including voice onset region(VOR), vowels and formants to distinguish British English and American English. By applying both linear neural network and neural network with nonlinear classifications and two hidden layers(NN2HL), the accuracy rate reaches 86.67%, which is very satisfying.